SOTA-Class TTS at Compact Scale
Python
Prepare Realistic 0.6B release docs and listening demos
chuarunyuana75cdbef546a
diff --git a/README.md b/README.mdindex 59c66c2..8edcf64 100644--- a/README.md+++ b/README.md@@ -1,502 +1,157 @@<div align="center">-<img src="assets/20260729-124515.jpeg" alt="Audio8" width="100%">+<img src="assets/20260729-124515.jpeg" alt="Audio8" width="760">-# Audio8_TTS Preview+<h1 align="center">Audio8 TTS</h1>-**A 0.6B-parameter multilingual text-to-speech model with zero-shot voice cloning.**+**Compact text-to-speech models for zero-shot voice cloning.**-[](https://github.com/Audio8-AI/Audio8_TTS)-[](https://audio8-ai.github.io/Audio8_TTS/)-[](https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b)-[](https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6B-ONNX-INT4)-[](LICENSE)+[](https://github.com/Edge0-AI/Audio8_TTS)+[](LICENSE)-中文文档: [README_zh.md](README_zh.md)+[English](README.md) · [中文](README_zh.md)-</div>--> 🚀 **How small a zero-shot cloning TTS can be?** Meet our latest release,-> **Audio8-TTS-0.1B**: a powerful, compact, and even portable text-to-speech-> model!->-> [](https://huggingface.co/Audio8/Audio8-TTS-Preview-0.1b)-> [](https://huggingface.co/Audio8/audio8-TTS-0.1B-ONNX-INT8)-> [](https://audio8-ai.github.io/Audio8_TTS/0.1B/)--This repository provides the audio8_tts Preview checkpoint, Hugging Face remote-code, inference tools, and an independent SFT pipeline for multilingual speech-generation and zero-shot voice cloning.--> **Preview status:** language coverage is intentionally limited in this-> release. Use the model primarily with the 11 recommended languages below.-> Multilingual coverage and Chinese dialect support will be expanded in later-> releases.--## Supported Languages--The Preview checkpoint performs best in the following languages:--| Language | Name |-|---|---|-| Cantonese | 粤语 |-| Chinese | 中文 |-| Dutch | 荷兰语 |-| English | 英语 |-| French | 法语 |-| German | 德语 |-| Italian | 意大利语 |-| Japanese | 日语 |-| Korean | 韩语 |-| Polish | 波兰语 |-| Spanish | 西班牙语 |--## Architecture--audio8_tts uses a DualAR architecture inspired by-[Fish Audio S2 Pro](https://github.com/fishaudio/fish-speech).--| Component | Configuration |-|---|---|-| Main model | 601,159,424 parameters, excluding the codec |-| Slow AR | 24 layers, width 896, 14 attention heads, 2 KV heads |-| Fast AR | 4 layers, width 896, 14 attention heads, 2 KV heads |-| Acoustic tokens | 10 codebooks, 4,096 entries per codebook |-| Codec | 44.1 kHz, 2,048 samples per model frame (~21.5 frames/s) |-| Context | Up to 2,048 packed text/audio positions |--The slow AR transformer predicts one semantic token for each audio frame. The-fast AR transformer then predicts the frame's codec codebooks, conditioned on-the slow hidden state and preceding codebooks. Static KV caches are used by-both branches during generation. The checkpoint also bundles its neural codec,-so reference encoding and waveform decoding require no separate model.--## Installation--Python 3.10 or newer and a CUDA-capable GPU are recommended.--```bash-python3 -m venv .venv-source .venv/bin/activate-pip install -r requirements.txt-```--Download the checkpoint from-[Hugging Face](https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b) and-place it in the repository's `model/` directory. The expected local checkpoint-path is `model/audio8_tts_0_6B_preview/`. All commands also accept a Hugging-Face model ID through `--model`.+[0.6B family](#family-06b) · [0.1B family](#family-01b)-## Inference--For best synthesis quality, keep each input within 150 characters. Longer text-may reduce generation quality; split it into shorter segments when needed.--### Zero-shot voice cloning--The reference transcript should match the spoken content in the reference-audio.--```bash-python audio8_tts_infer.py \- --text "Welcome to audio8_tts." \- --reference-audio examples/reference.wav \- --reference-text "Transcript of the reference recording." \- --output outputs/clone.wav-```--### Generation without a reference--```bash-python audio8_tts_infer.py \- --text "This utterance does not use a reference voice." \- --output outputs/no_reference.wav-```+</div>-### Batch inference+## Releases-Each line in the input manifest is an independent JSON object. Relative audio-paths are resolved from the manifest directory.+<a id="family-06b"></a>-```json-{"id":"sample_001","text":"Target text","reference_audio":"audio/ref.wav","reference_text":"Reference transcript"}-{"id":"sample_002","text":"Text without a reference voice"}-```+<details open>+<summary><strong>0.6B family</strong></summary>-```bash-python audio8_tts_infer.py \- --input-jsonl data/prompts.jsonl \- --output-dir outputs/batch \- --batch-size 2-```+<div align="center">-The batch command writes `manifest.jsonl` and `failures.jsonl`. Existing WAV-files are skipped unless `--overwrite` is passed. See-`python audio8_tts_infer.py --help` for sampling and code-saving options.+<img src="assets/readme-06b.svg" alt="Audio8-TTS-Realistic — More realistic speech. More languages. Still 0.6B." width="960">-## CPU ONNX Runtime+[](https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b)+[](https://edge0-ai.github.io/Audio8_TTS/realistic-0.6b/)+[](https://arxiv.org/abs/2610.14776)-[`onnx_runtime/`](onnx_runtime/) provides a standalone CPU deployment using-weight-only INT4 Slow/Fast AR models, FP16 activations and KV caches, and an-FP16 codec. It includes CLI inference, a local web and HTTP service, streaming-PCM output, and reference-voice registration without PyTorch or Transformers.+</div>-The online sessions use about 1 GiB of memory in the tested Apple M2 setup.-During voice registration, the online sessions are released before the codec-encoder is loaded to keep peak memory low.+**Audio8-TTS-Realistic-0.6B** builds on Preview with substantially improved speech realism, retaining the same compact 0.6B architecture and zero-shot voice cloning. It supports **23 languages**, adding **12 languages compared with Preview**.-Download the ONNX model from-[Audio8-TTS-Preview-0.6B-ONNX-INT4](https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6B-ONNX-INT4)-and follow the [ONNX Runtime guide](onnx_runtime/README.md).+Audio8 is also participating in [TTS Arena](https://tts-agi-tts-arena-v2.hf.space/), where it has received encouraging ratings from listeners. We are grateful to everyone who has listened and voted.-## SGLang Omni Serving+| Release | Focus and languages | Model | Listen | Guide |+|---|---|---|---|---|+| **Realistic 0.6B** | Realism and prosody · 23 languages | [HF](https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b) | [Demo](https://edge0-ai.github.io/Audio8_TTS/realistic-0.6b/) | [Details](guides/models/realistic-0.6b.md) |-The adapter in [`sglang_omni/`](sglang_omni/) provides an OpenAI-compatible-service with SGLang paged attention, dynamic batching, a fixed KV cache for the-fast codebook decoder, reference-audio encoding, and waveform decoding. It is-installed as an independent `audio8_tts` model plugin and does not overwrite-SGLang Omni core files.+**ONNX:** There is currently no Realistic ONNX export.-### Compatibility+<details>+<summary><strong>0.6B details: languages, quick start, and evaluation</strong></summary>-The adapter uses internal SGLang Omni interfaces, so deploy it with the tested-revision instead of the latest `main` branch.+#### Realistic 0.6B language support-| Dependency | Tested version |+| Coverage | Languages ||---|---|-| SGLang Omni | `68a572348837f7b004857b4b07993c20ade4c017` (`0.1.0`) |-| SGLang | `0.5.8` |-| PyTorch | `2.9.1+cu128` |-| Transformers | `4.57.1` |-| Precision | BF16 |--### Audio8-TTS-0.1B (Falcon-H1 hybrid slow backbone)--`Audio8-TTS-Preview-0.1b` replaces the pure-attention slow AR backbone with a-Falcon-H1 hybrid (Mamba 2 SSM + attention). The adapter detects this-automatically from `config.json` (`slow_backbone: falcon_h1` or a `mamba_d_ssm`-field) and switches to the eager hybrid path: the slow backbone runs with one-`FalconHybridMambaAttentionDynamicCache` per request, while the fast codebook-AR, semantic sampling and vocoder are shared with the 0.6B path. The 0.6B-path is unchanged.--```bash-export MODEL=/models/Audio8-TTS-Preview-0.1b-export CONFIG=./sglang_omni/configs/audio8_tts_0_1b.yaml-export MODEL_NAME=audio8/tts-0.1b--CUDA_VISIBLE_DEVICES=0 \-SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \-MODEL="${MODEL}" \-CONFIG="${CONFIG}" \-MODEL_NAME="${MODEL_NAME}" \-./sglang_omni/scripts/run_server.sh-```--Because the Mamba backbone keeps its own hybrid caches and does not consume-SGLang KV pages, the adapter disables CUDA graphs and caps the static memory-fraction for this model. The slow backbone is bit-exact with the Transformers-Falcon-H1 implementation for both prefill and cached decode.+| 11 languages supported in Preview | Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Spanish |+| 12 additions in Realistic | Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish, Vietnamese |-### Performance+> **Note:** Quality varies by language; benchmark coverage differs from the supported-language list. [Per-language results](guides/models/realistic-0.6b.md#multilingual-detail).-Warm single-stream latency was measured on one NVIDIA H20 with BF16 weights,-CUDA Graph, greedy decoding, and 128 generated frames. The output WAV was-5.85-5.94 seconds long; cold start and compilation time were excluded. Lower-RTF is better.+#### Realistic 0.6B quick start-| SGLang Omni adapter | Warm p50 latency | RTF |-|---|---:|---:|-| Current implementation | **0.691 s** | **0.116** |--See the [SGLang Omni implementation and evaluation report](sglang_omni/OPTIMIZATION_REPORT.md)-for the configuration, implementation details, and validation results.--### Install--Run these commands from the Audio8 TTS repository root. The example uses-Python 3.12 and [`uv`](https://docs.astral.sh/uv/).+Use Python 3.10+ and a CUDA-capable GPU. Run from the repository root. The explicit model ID selects Realistic independently of the scripts' existing Preview default.```bash-export SGLANG_OMNI_ROOT=/opt/sglang-omni-export MODEL=/models/Audio8-TTS-Preview-0.6b--git clone https://github.com/sgl-project/sglang-omni.git "${SGLANG_OMNI_ROOT}"-git -C "${SGLANG_OMNI_ROOT}" checkout 68a572348837f7b004857b4b07993c20ade4c017--uv venv .venv-sglang --python 3.12-source .venv-sglang/bin/activate-uv pip install -v -e "${SGLANG_OMNI_ROOT}"+git clone https://github.com/Edge0-AI/Audio8_TTS.git+cd Audio8_TTS+python3 -m venv .venv+source .venv/bin/activate+pip install -r requirements.txt-hf download AutoArk-AI/Audio8-TTS-Preview-0.6b --local-dir "${MODEL}"-./sglang_omni/scripts/install_adapter.sh "${SGLANG_OMNI_ROOT}"-python3 ./sglang_omni/scripts/verify_install.py --model-path "${MODEL}"+python audio8_tts_infer.py \+ --model Edge0/Audio8-TTS-Realistic-0.6b \+ --text "Welcome to Audio8 TTS. It is wonderful to hear from you." \+ --output outputs/realistic.wav```-For an existing wheel or site-packages installation, resolve the package-directory and install the adapter there:+For voice cloning, add `--reference-audio reference.wav --reference-text "The exact transcript of the recording."`.-```bash-SGLANG_OMNI_PACKAGE="$(python3 -c 'import importlib.util, pathlib; s=importlib.util.find_spec("sglang_omni"); assert s and s.origin; print(pathlib.Path(s.origin).parent)')"-./sglang_omni/scripts/install_adapter.sh "${SGLANG_OMNI_PACKAGE}"-```+[Installation, cloning and batch inference](guides/inference.md) · [Supervised fine-tuning](guides/training.md)-### Start the service+#### 0.6B evaluation summary-```bash-CUDA_VISIBLE_DEVICES=0 \-SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \-MODEL="${MODEL}" \-AUDIO8_TTS_ENABLE_TORCH_COMPILE=1 \-HOST=0.0.0.0 \-PORT=8010 \-./sglang_omni/scripts/run_server.sh-```+Error rates are percentages (lower is better). SIM uses a 0–1 scale (higher is better). Neither directly measures speech realism.-The default `fa3` attention backend is intended for Hopper GPUs such as H20-and H100. Consumer Blackwell GPUs such as RTX 5090 report compute capability-`(12, 0)` and have no FA3 kernel image, so the adapter detects them and selects-FlashInfer for the SGLang slow-AR path automatically; the short fixed-cache-fast head then uses PyTorch SDPA. No configuration is required.+**Audio8-TTS-Preview** and **Audio8-TTS-Realistic**, both **0.6B**:-Setting the variable explicitly still overrides the detection on any GPU:+| Benchmark | Coverage | Preview error % ↓ | Realistic error % ↓ | Preview SIM ↑ | Realistic SIM ↑ |+|---|---|---:|---:|---:|---:|+| Seed-TTS | English / Chinese | 1.23 | 1.25 | 0.682 | 0.704 |+| CV3 | 8 shared languages | 4.80 | 4.56 | — | 0.663 |+| CV3 | 9 core languages | — | 4.86 | — | 0.661 |+| MiniMax | 24 evaluated languages | — | 5.89 | — | 0.754 |-```bash-AUDIO8_TTS_ATTENTION_BACKEND=flashinfer \-CUDA_VISIBLE_DEVICES=0 \-SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \-MODEL="${MODEL}" \-./sglang_omni/scripts/run_server.sh-```+> **Note:** Values are unweighted language macro averages. Preview results are calculated from its [published model card](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b#evaluation), with SIM converted from percentages to a 0–1 scale. The two releases were evaluated separately. A dash means the corresponding result was not reported on Preview's card.+>+> Seed-TTS averages English WER and Chinese CER, excluding hard Chinese. The shared CV3 subset is Chinese, English, Japanese, Korean, German, French, Spanish, and Italian; the 9-language set also includes Russian. CV3 averages exclude hard subsets and the 13 additional languages. MiniMax averages all 24 evaluated languages, including languages outside the supported list.-The defaults use model name `audio8/tts-0.6b`, BF16, one GPU, a `0.2` static-memory fraction, and up to 32 running requests. The main runtime controls are-`MODEL_NAME`, `AUDIO8_TTS_MEM_FRACTION_STATIC`,-`AUDIO8_TTS_MAX_RUNNING_REQUESTS`, `AUDIO8_TTS_CHUNKED_PREFILL_SIZE`, and-`AUDIO8_TTS_DISABLE_CUDA_GRAPH`. When Torch compilation is enabled, the adapter-uses SGLang's native batch-size policy. Set `AUDIO8_TTS_TORCH_COMPILE_MAX_BS`-only when an explicit compile limit is needed. `AUDIO8_TTS_ATTENTION_BACKEND`-defaults to `fa3`, except on GPUs with no FA3 kernel image such as consumer-Blackwell, where it defaults to `flashinfer`; set it explicitly to override. Set-`SGLANG_OMNI_SITE_PACKAGES` when the runtime dependencies are installed in a-separate site-packages directory.+[Full Realistic evaluation and comparison notes](guides/models/realistic-0.6b.md#evaluation) · [Original Preview evaluation](guides/models/preview-0.6b.md#evaluation)-### Troubleshooting+</details>-- Install the distribution package that provides `libnuma` if `sgl_kernel`- fails to import (for example, `numactl` or `libnuma1`).-- Put the CUDA toolkit `bin` directory on `PATH`. If `deep_gemm` cannot find- `nvcc` during JIT compilation, also set `CUDA_PATH` to the toolkit root.-- Keep Transformers on the supported 4.x range (`>=4.57.0,<5`). Transformers- 5.x can produce invalid all-zero codes for this custom-code model.+<details>+<summary><strong>Previous release · Preview 0.6B</strong></summary>-### Call the API+Realistic builds on Preview. The original checkpoint, **11-language** coverage, evaluation, and listening page remain available for Preview users.-Generate speech without a reference:+| Release | Model | Listen | Guide | Deployment |+|---|---|---|---|---|+| Preview 0.6B | [HF](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b) | [Demo](https://edge0-ai.github.io/Audio8_TTS/preview-0.6b/) | [Details](guides/models/preview-0.6b.md) | [ONNX INT4](onnx_runtime/README.md) |-```bash-curl -sS --fail-with-body \- -H 'Content-Type: application/json' \- -d '{- "model": "audio8/tts-0.6b",- "input": "Hello from Audio8 TTS.",- "response_format": "wav",- "max_new_tokens": 256,- "temperature": 0.8,- "top_p": 0.95,- "top_k": 50- }' \- http://127.0.0.1:8010/v1/audio/speech \- -o audio8.wav-```+</details>-Generate speech with one reference voice:+</details>-```bash-curl -sS --fail-with-body \- -H 'Content-Type: application/json' \- -d '{- "model": "audio8/tts-0.6b",- "input": "This sentence uses the reference voice.",- "response_format": "wav",- "temperature": 0.8,- "top_p": 0.95,- "top_k": 50,- "references": [{- "audio_path": "/data/reference.wav",- "text": "The exact transcript of the reference recording."- }]- }' \- http://127.0.0.1:8010/v1/audio/speech \- -o audio8_clone.wav-```+----The reference path must be visible inside the service environment. The current-adapter supports TP=1 and one reference per request. Set `"stream": true` to-receive SSE audio chunks as they are generated. For the lowest overhead, use-`"response_format": "pcm"`; each event contains Base64-encoded audio in-`audio.data`. Streaming defaults to 12 codec frames per chunk, with 128 frames-of decoder context and a one-frame boundary guard. Override these values with-`AUDIO8_TTS_STREAM_CHUNK_FRAMES`, `AUDIO8_TTS_STREAM_CONTEXT_FRAMES`, and-`AUDIO8_TTS_STREAM_GUARD_FRAMES`.+<a id="family-01b"></a>-Streaming is a server-side opt-in. Start the service with-`AUDIO8_TTS_STREAM_ENABLED=1` to enable SSE streaming; the default is off, and-requests are answered with the complete audio after generation finishes-(a request-level `"stream": true` is then ignored). Non-streaming remains the-default for maximum throughput and stable memory usage. `response_format` of-`codes`/`codec`/`npy` always returns codec codes without streaming.+<details open>+<summary><strong>0.1B family</strong></summary>-Run the smoke test to verify a deployment:+<div align="center">-```bash-BASE_URL=http://127.0.0.1:8010 ./sglang_omni/scripts/smoke_test.sh-python3 ./sglang_omni/scripts/stream_smoke_test.py \- --base-url http://127.0.0.1:8010 \- --output /tmp/audio8_stream.wav-```+<img src="assets/readme-01b.svg" alt="Audio8-TTS-0.1B — How small a zero-shot cloning TTS can be?" width="960">-To build the adapter into an existing image, append-[`sglang_omni/Dockerfile.snippet`](sglang_omni/Dockerfile.snippet) after the-SGLang Omni package and its Python dependencies are installed.+[](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b)+[](https://huggingface.co/Edge0/audio8-TTS-0.1B-ONNX-INT8)+[](https://edge0-ai.github.io/Audio8_TTS/0.1B/)-## Supervised Fine-tuning+</div>-Install the training dependencies first:+Meet **Audio8-TTS-0.1B**: a compact, portable text-to-speech model with zero-shot voice cloning.-```bash-pip install -r requirements-train.txt-```+Designed for a smaller deployment footprint, with **Chinese and English as its primary languages**. The dedicated ONNX INT8 runtime supports CPU inference, streaming, and local web/HTTP serving.-### 1. Create a raw manifest+| Release | Focus and languages | Model | Listen | Guide |+|---|---|---|---|---|+| Preview 0.1B | Compact hybrid · primarily Chinese / English | [HF](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b) | [Demo](https://edge0-ai.github.io/Audio8_TTS/0.1B/) | [Details](guides/models/preview-0.1b.md) |-The target `audio` field is required. `reference_audio` and `reference_text`-are optional, but must be provided together.+[0.1B quick start](guides/models/preview-0.1b.md#inference) · [CPU / ONNX runtime](onnx_runtime_0_1b_int8/README.md) · [0.1B deployment](guides/models/preview-0.1b.md#deployment) · [Model license](guides/models/preview-0.1b.md#license)-```json-{"id":"utt_001","text":"Target transcript","audio":"audio/target.wav","reference_audio":"audio/reference.wav","reference_text":"Reference transcript"}-{"id":"utt_002","text":"Another transcript","audio":"audio/another.wav"}-```+</details>-### 2. Precompute codec indices+----```bash-python audio8_tts_prepare.py \- --input-jsonl data/train.jsonl \- --output-jsonl prepared_data/train.jsonl \- --batch-size 4-```+## Shared tools-The prepared manifest points to validated `[10, T]` NumPy arrays using paths-relative to the prepared manifest. Existing valid arrays are reused unless-`--overwrite` is passed.+[Installation and inference](guides/inference.md) · [Supervised fine-tuning](guides/training.md)-### 3. Train+The shared tools accept an explicit model ID. Use the quick-start and deployment instructions for your chosen release.-Single GPU:+## SGLang Omni Serving-```bash-TRAIN_JSONL=prepared_data/train.jsonl \-NPROC_PER_NODE=1 \-bash audio8_tts_sft.sh-```+SGLang Omni provides paged attention, dynamic batching, and an OpenAI-compatible speech API. The 0.6B family and the 0.1B hybrid model have their own configurations. Existing performance results remain paired with their original model, hardware, and test conditions in the deployment guide.-Eight GPUs on one node:+[Deployment matrix, installation and server configuration](guides/deployment.md) · [API, streaming and troubleshooting](guides/deployment.md#sglang-omni-serving)-```bash-TRAIN_JSONL=prepared_data/train.jsonl \-NPROC_PER_NODE=8 \-BATCH_SIZE=2 \-GRADIENT_ACCUMULATION_STEPS=8 \-bash audio8_tts_sft.sh-```+## License and acknowledgements-For multi-node training, set `NNODES`, `NODE_RANK`, `MASTER_ADDR`, and-`MASTER_PORT` on each node. Common hyperparameters and output paths can be-overridden through the environment variables in `audio8_tts_sft.sh`; additional-Transformers arguments may be appended to the command.--SFT optimizes both the slow semantic/EOS objective and the fast codebook-teacher-forcing objective. Set `FREEZE_SLOW_AR=true` or `FREEZE_FAST_AR=true`-when adapting only one branch. The exported directory remains loadable with-standard `AutoModel` and `AutoProcessor` APIs using `trust_remote_code=True`.--## Evaluation--Audio8 TTS Preview is the smallest model in this comparison at just **0.6B-parameters**. Despite using only a fraction of the parameters of the other-systems, it delivers results in the first tier of industry-leading SOTA TTS-models on the benchmarks below. In particular, it achieves the best English-WER and competitive Chinese CER on Seed-TTS, while remaining competitive-across the CV3 multilingual evaluation.--Lower WER/CER is better; higher SIM is better. Seed-TTS similarity values are-shown as percentages.--### Seed-TTS--| Model | Parameters | EN WER / SIM | ZH CER / SIM | Hard ZH CER / SIM |-|---|---:|---:|---:|---:|-| **Audio8 TTS Preview** | **0.6B** | **1.506** / 63.2 | 0.950 / 73.1 | 11.510 / 68.7 |-| Fish S2 Pro | 4.6B | 1.607 / 64.6 | 1.038 / 73.8 | 10.149 / 70.1 |-| Higgs Audio v2 | 4.7B | 1.524 / 66.4 | **0.806** / 72.1 | 10.622 / 69.3 |-| CosyVoice3-1.5B | 1.5B | 2.22 / 72.0 | 1.12 / 78.1 | **5.83** / **75.8** |-| MOSS-TTS | 8.5B | 1.85 / 73.4 | 1.20 / 78.8 | - |-| VoxCPM2 | 2.3B | 1.84 / **75.3** | 0.97 / **79.5** | 8.13 / 75.3 |----### CV3 multilingual error rate--| Model | Parameters | zh | en | hard-zh | hard-en | ja | ko | de | es | fr | it | ru |-|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|-| **Audio8 TTS Preview** | **0.6B** | **3.205** | **3.128** | 10.535 | 5.997 | 7.205 | 4.223 | 3.447 | 3.641 | 8.790 | 4.790 | - |-| Fish S2 Pro | 4.6B | 3.600 | 3.493 | 10.588 | 7.349 | 5.139 | **4.111** | 3.605 | 2.972 | **8.600** | 4.229 | **4.702** |-| Higgs Audio v2 | 4.7B | 3.378 | 3.404 | 10.424 | **5.754** | **4.742** | 4.260 | **3.300** | **2.929** | 9.425 | **3.555** | 5.423 |-| CosyVoice3-1.5B | 1.5B | 3.91 | 4.99 | 9.77 | 10.55 | 7.57 | 5.69 | 6.43 | 4.47 | 11.8 | 10.5 | 6.64 |-| VoxCPM2 | 2.3B | 3.65 | 5.00 | **8.55** | 8.48 | 5.96 | 5.69 | 4.77 | 3.80 | 9.85 | 4.25 | 5.21 |----Parameter counts are calculated directly from the released weight tensors.-MOSS-TTS contains 8,489,841,664 parameters. VoxCPM2's main model contains-2,290,004,544 parameters; the separate AudioVAE is not included in the-parameter comparison.--Fish S2 Pro was reevaluated because its official evaluation uses its own-normalizer. Higgs Audio v2 was evaluated locally because concrete values were-unavailable. All other baseline values were collected from their official-reports through the [VoxCPM repository](https://github.com/OpenBMB/VoxCPM).--Different normalizers and evaluators make cross-project values reference-comparisons rather than a strictly matched ranking. Evaluation coverage does-not expand the Preview's supported-language claim beyond the 11 languages-listed above.--## Limitations and Responsible Use--- This is a Preview checkpoint with limited multilingual and dialect coverage.-- Very long, noisy, or inaccurate reference clips can reduce stability and- speaker similarity.-- Generated speech can be misused for impersonation or misinformation. Obtain- consent before cloning a voice and clearly disclose synthetic audio where- appropriate.-- Test the model for accuracy, safety, and legal compliance before deployment.--## License and Acknowledgements--Code and model weights in this repository are released under the-[Apache License 2.0](LICENSE). See [NOTICE](NOTICE) for attribution details.--We thank the Fish Audio team for publishing the DualAR architecture used in-Fish S2 Pro.--## Star History--<a href="https://www.star-history.com/?type=date&repos=Audio8-AI%2FAudio8_TTS">- <picture>- <source media="(prefers-color-scheme: dark)" srcset="https://api.star-history.com/chart?repos=Audio8-AI/Audio8_TTS&type=date&theme=dark&legend=top-left&sealed_token=ShFu9kcwBvymYQ4SjQ_NhkplHrefNRbYVYCiBIvIxnaBLKbEQ1cjHQBs2kZm7K5LNMWpU13JxWgA6zpHvmwh49FokyJ26axmq-0gG8b68Q8IJCyUDZW1jQ" />- <source media="(prefers-color-scheme: light)" srcset="https://api.star-history.com/chart?repos=Audio8-AI/Audio8_TTS&type=date&legend=top-left&sealed_token=ShFu9kcwBvymYQ4SjQ_NhkplHrefNRbYVYCiBIvIxnaBLKbEQ1cjHQBs2kZm7K5LNMWpU13JxWgA6zpHvmwh49FokyJ26axmq-0gG8b68Q8IJCyUDZW1jQ" />- <img alt="Star History Chart" src="https://api.star-history.com/chart?repos=Audio8-AI/Audio8_TTS&type=date&legend=top-left&sealed_token=ShFu9kcwBvymYQ4SjQ_NhkplHrefNRbYVYCiBIvIxnaBLKbEQ1cjHQBs2kZm7K5LNMWpU13JxWgA6zpHvmwh49FokyJ26axmq-0gG8b68Q8IJCyUDZW1jQ" />- </picture>-</a>+The repository code and both 0.6B releases use [Apache License 2.0](LICENSE). The 0.1B checkpoint has its own [model license](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b#license); consult the license packaged with each model/export. See [NOTICE](NOTICE) for attribution. We thank the Fish Audio team for the DualAR architecture used in Fish Audio S2 Pro.diff --git a/README_zh.md b/README_zh.mdindex 9c6086d..9ebb6db 100644--- a/README_zh.md+++ b/README_zh.md@@ -1,449 +1,155 @@<div align="center">-<img src="assets/20260729-124515.jpeg" alt="Audio8" width="100%">+<img src="assets/20260729-124515.jpeg" alt="Audio8" width="760">-# Audio8_TTS Preview+<h1 align="center">Audio8 TTS</h1>-**支持多语言语音生成和零样本音色克隆的 0.6B 参数文本转语音模型。**+**轻量语音合成模型系列,支持零样本音色克隆。**-[](https://github.com/Audio8-AI/Audio8_TTS)-[](https://audio8-ai.github.io/Audio8_TTS/)-[](https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b)-[](https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6B-ONNX-INT4)-[](LICENSE)+[](https://github.com/Edge0-AI/Audio8_TTS)+[](LICENSE)-English: [README.md](README.md)+[English](README.md) · [中文](README_zh.md)-</div>--本仓库提供 audio8_tts Preview 版模型、Hugging Face remote code、推理工具和独立的 SFT-训练流程,支持多语言语音生成和零样本音色克隆。--> **Preview 说明:** 当前版本的语言覆盖仍然有限,建议优先在下列 11 种语言中使用。后续版本将-> 持续补齐多语言和中文方言能力。--## 支持语言--Preview 模型当前表现较好、推荐使用的语言如下:--| Language | 语言 |-|---|---|-| Cantonese | 粤语 |-| Chinese | 中文 |-| Dutch | 荷兰语 |-| English | 英语 |-| French | 法语 |-| German | 德语 |-| Italian | 意大利语 |-| Japanese | 日语 |-| Korean | 韩语 |-| Polish | 波兰语 |-| Spanish | 西班牙语 |--## 模型结构--audio8_tts 沿用了 [Fish Audio S2 Pro](https://github.com/fishaudio/fish-speech)-的 DualAR 架构思路。--| 组件 | 配置 |-|---|---|-| 主模型 | 601,159,424 参数,不包含 codec |-| Slow AR | 24 层、896 维、14 个 attention head、2 个 KV head |-| Fast AR | 4 层、896 维、14 个 attention head、2 个 KV head |-| 声学 token | 10 个 codebook,每个 codebook 包含 4,096 个条目 |-| Codec | 44.1 kHz,每个模型帧 2,048 个采样点,约 21.5 帧/秒 |-| 上下文 | 最多 2,048 个打包后的文本/音频位置 |--Slow AR 每个音频帧生成一个 semantic token;Fast AR 根据 slow hidden state 和当前帧已经生成的-codebook,继续生成该帧的十个 codec codebook。推理时 slow AR 和 fast AR 都使用静态 KV cache。-Checkpoint 内置神经音频 codec,不需要额外下载参考音频编码器或波形解码器。--## 安装--推荐使用 Python 3.10 或更高版本以及支持 CUDA 的 GPU。--```bash-python3 -m venv .venv-source .venv/bin/activate-pip install -r requirements.txt-```--请从 [Hugging Face](https://huggingface.co/AutoArk-AI/Audio8-TTS-Preview-0.6b)-下载模型,并将其放入仓库的 `model/` 文件夹中。默认本地模型路径为-`model/audio8_tts_0_6B_preview/`。所有命令也可以通过 `--model` 接收 Hugging Face 模型 ID。--## 推理--为获得更好的合成效果,建议单次输入不超过 150 个字。文本过长可能影响生成质量,请根据-需要拆分为较短的片段。--### 零样本音色克隆--参考文本应当与参考音频中实际说出的内容一致。--```bash-python audio8_tts_infer.py \- --text "欢迎使用 audio8_tts。" \- --reference-audio examples/reference.wav \- --reference-text "参考录音对应的准确文本。" \- --output outputs/clone.wav-```--### 无参考音频生成+[0.6B family](#family-06b) · [0.1B family](#family-01b)-```bash-python audio8_tts_infer.py \- --text "这条语音不使用参考音色。" \- --output outputs/no_reference.wav-```+</div>-### 批量推理+## 模型版本-输入文件每行是一个独立 JSON 对象。相对音频路径相对于 JSONL 文件所在目录解析。+<a id="family-06b"></a>-```json-{"id":"sample_001","text":"目标文本","reference_audio":"audio/ref.wav","reference_text":"参考文本"}-{"id":"sample_002","text":"不使用参考音色的文本"}-```+<details open>+<summary><strong>0.6B family</strong></summary>-```bash-python audio8_tts_infer.py \- --input-jsonl data/prompts.jsonl \- --output-dir outputs/batch \- --batch-size 2-```+<div align="center">-批量命令会生成 `manifest.jsonl` 和 `failures.jsonl`。除非传入 `--overwrite`,否则已有 WAV 会-被跳过。采样和 codec codes 保存参数可通过 `python audio8_tts_infer.py --help` 查看。+<img src="assets/readme-06b-zh.svg" alt="Audio8-TTS-Realistic — 更真实的语音。更广的语言覆盖。依然只有 0.6B。" width="960">-## CPU ONNX Runtime+[](https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b)+[](https://edge0-ai.github.io/Audio8_TTS/realistic-0.6b/)+[](https://arxiv.org/abs/2610.14776)-[`onnx_runtime/`](onnx_runtime/) 提供独立的 CPU 部署方案:Slow/Fast AR 使用-weight-only INT4,activation 和 KV cache 使用 FP16,codec 使用 FP16。它包含-命令行推理、本地网页与 HTTP 服务、流式 PCM 输出和参考音频音色注册,运行时不依赖-PyTorch 或 Transformers。+</div>-在 Apple M2 测试环境中,在线模型 session 约占 1 GiB 内存。注册音色前会先释放-在线 session,再单独加载 codec encoder,从而控制峰值内存。+**Audio8-TTS-Realistic-0.6B** 在 Preview 基础上显著提升语音拟真感,保留相同的 0.6B 架构和零样本音色克隆能力。支持 **23 种语言**,相较 Preview 新增 **12 种**。-从-[Audio8-TTS-Preview-0.6B-ONNX-INT4](https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6B-ONNX-INT4)-下载模型,并参阅 [ONNX Runtime 中文指南](onnx_runtime/README_zh.md)。+Audio8 也参与了 [TTS Arena](https://tts-agi-tts-arena-v2.hf.space/),目前收到的听众评分令人鼓舞。感谢每一位参与试听和投票的朋友。-## SGLang Omni 服务部署+| 版本 | 定位与语言 | 模型 | 试听 | 版本指南 |+|---|---|---|---|---|+| **Realistic 0.6B** | 拟真感与韵律 · 23 种语言 | [HF](https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b) | [Demo](https://edge0-ai.github.io/Audio8_TTS/realistic-0.6b/) | [说明](guides/models/realistic-0.6b_zh.md) |-[`sglang_omni/`](sglang_omni/) 中的适配器提供兼容 OpenAI API 的推理服务,包含 SGLang-paged attention、动态 batching、Fast AR 固定 KV cache、参考音频编码和 waveform 解码。-它以独立的 `audio8_tts` 模型插件安装,不会覆盖 SGLang Omni 核心文件。+**ONNX:** Realistic 目前暂无 ONNX 制品。-### 兼容版本+<details>+<summary><strong>0.6B 详情:语言支持、快速开始与评测</strong></summary>-适配器使用了 SGLang Omni 的内部接口,因此部署时应固定到以下已验证版本,不要直接使用最新-`main` 分支。+#### Realistic 0.6B 的语言支持-| 依赖 | 已验证版本 |+| 范围 | 语言 ||---|---|-| SGLang Omni | `68a572348837f7b004857b4b07993c20ade4c017`(`0.1.0`) |-| SGLang | `0.5.8` |-| PyTorch | `2.9.1+cu128` |-| Transformers | `4.57.1` |-| 精度 | BF16 |--### Audio8-TTS-0.1B(Falcon-H1 混合慢速主干)--`Audio8-TTS-Preview-0.1b` 将纯注意力的慢速 AR 主干替换为 Falcon-H1 混合结构-(Mamba 2 SSM + 注意力)。适配器会自动从 `config.json` 检测(`slow_backbone:-falcon_h1` 或存在 `mamba_d_ssm` 字段)并切换到 eager 混合路径:慢速主干为每个-请求维护一个 `FalconHybridMambaAttentionDynamicCache`,而快速码本 AR、语义采样-与声码器与 0.6B 路径完全共用,0.6B 路径不受影响。--```bash-export MODEL=/models/Audio8-TTS-Preview-0.1b-export CONFIG=./sglang_omni/configs/audio8_tts_0_1b.yaml-export MODEL_NAME=audio8/tts-0.1b--CUDA_VISIBLE_DEVICES=0 \-SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \-MODEL="${MODEL}" \-CONFIG="${CONFIG}" \-MODEL_NAME="${MODEL_NAME}" \-./sglang_omni/scripts/run_server.sh-```+| Preview 已支持的 11 种 | 粤语、中文、荷兰语、英语、法语、德语、意大利语、日语、韩语、波兰语、西班牙语 |+| Realistic 新增的 12 种 | 阿拉伯语、芬兰语、印地语、印度尼西亚语、马来语、挪威语、葡萄牙语、俄语、瑞典语、泰语、土耳其语、越南语 |-由于 Mamba 主干自行维护混合缓存、不占用 SGLang KV 页,适配器会为该模型关闭-CUDA Graph 并限制静态显存比例。慢速主干在 prefill 与增量 decode 上与-Transformers 的 Falcon-H1 实现逐位一致。+> **说明:** 语言质量存在差异;评测覆盖和正式支持清单并不相同。 [分语言评测](guides/models/realistic-0.6b_zh.md#多语言详细结果)。-### 性能+#### Realistic 0.6B 快速开始-单流 warm latency 在单张 NVIDIA H20 上测试,使用 BF16、CUDA Graph、greedy decoding,-并生成 128 帧。输出 WAV 时长为 5.85-5.94 秒,不包含冷启动和编译时间。RTF 越低越好。--| SGLang Omni 适配器 | Warm p50 latency | RTF |-|---|---:|---:|-| 当前实现 | **0.691 s** | **0.116** |--配置、实现细节和验证结果请参阅-[SGLang Omni 实现与评测报告](sglang_omni/OPTIMIZATION_REPORT.md)。--### 安装--在 Audio8 TTS 仓库根目录执行以下命令。示例使用 Python 3.12 和-[`uv`](https://docs.astral.sh/uv/)。+推荐 Python 3.10+ 和 CUDA GPU;在仓库根目录执行。模型 ID 显式指定,因此不会误用脚本保留的 Preview 默认路径。```bash-export SGLANG_OMNI_ROOT=/opt/sglang-omni-export MODEL=/models/Audio8-TTS-Preview-0.6b--git clone https://github.com/sgl-project/sglang-omni.git "${SGLANG_OMNI_ROOT}"-git -C "${SGLANG_OMNI_ROOT}" checkout 68a572348837f7b004857b4b07993c20ade4c017--uv venv .venv-sglang --python 3.12-source .venv-sglang/bin/activate-uv pip install -v -e "${SGLANG_OMNI_ROOT}"--hf download AutoArk-AI/Audio8-TTS-Preview-0.6b --local-dir "${MODEL}"-./sglang_omni/scripts/install_adapter.sh "${SGLANG_OMNI_ROOT}"-python3 ./sglang_omni/scripts/verify_install.py --model-path "${MODEL}"-```--如果 SGLang Omni 已通过 wheel 或 site-packages 安装,可以定位其包目录后安装适配器:--```bash-SGLANG_OMNI_PACKAGE="$(python3 -c 'import importlib.util, pathlib; s=importlib.util.find_spec("sglang_omni"); assert s and s.origin; print(pathlib.Path(s.origin).parent)')"-./sglang_omni/scripts/install_adapter.sh "${SGLANG_OMNI_PACKAGE}"-```--### 启动服务--```bash-CUDA_VISIBLE_DEVICES=0 \-SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \-MODEL="${MODEL}" \-AUDIO8_TTS_ENABLE_TORCH_COMPILE=1 \-HOST=0.0.0.0 \-PORT=8010 \-./sglang_omni/scripts/run_server.sh-```--默认的 `fa3` attention backend 适用于 H20、H100 等 Hopper GPU。RTX 5090 等消费级-Blackwell GPU 的计算能力为 `(12, 0)`,没有对应的 FA3 kernel image,适配器会自动识别-并为 SGLang Slow AR 路径选择 FlashInfer;此时短序列固定 KV cache 的 Fast head 会使用-PyTorch SDPA。无需额外配置。--在任意 GPU 上显式设置该环境变量仍会覆盖自动检测结果:--```bash-AUDIO8_TTS_ATTENTION_BACKEND=flashinfer \-CUDA_VISIBLE_DEVICES=0 \-SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \-MODEL="${MODEL}" \-./sglang_omni/scripts/run_server.sh-```--默认配置使用模型名 `audio8/tts-0.6b`、BF16、单卡、`0.2` 静态显存比例和最多 32 个并发-请求。主要运行参数包括 `MODEL_NAME`、`AUDIO8_TTS_MEM_FRACTION_STATIC`、-`AUDIO8_TTS_MAX_RUNNING_REQUESTS`、`AUDIO8_TTS_CHUNKED_PREFILL_SIZE` 和-`AUDIO8_TTS_DISABLE_CUDA_GRAPH`。开启 Torch compile 后,适配器沿用 SGLang 原生的 batch-策略;只有需要明确限制 compile 上限时才设置 `AUDIO8_TTS_TORCH_COMPILE_MAX_BS`。-`AUDIO8_TTS_ATTENTION_BACKEND` 默认为 `fa3`;在没有 FA3 kernel image 的 GPU(如消费级-Blackwell)上默认为 `flashinfer`,显式设置可覆盖。如果运行依赖安装在单独的-site-packages 目录中,请设置-`SGLANG_OMNI_SITE_PACKAGES`。--### 常见问题--- 如果 `sgl_kernel` 导入失败,请安装系统中提供 `libnuma` 的软件包,例如 `numactl` 或- `libnuma1`。-- CUDA toolkit 的 `bin` 目录需要加入 `PATH`。如果 `deep_gemm` JIT 编译时找不到- `nvcc`,还需要将 `CUDA_PATH` 指向 CUDA toolkit 根目录。-- Transformers 应保持在支持的 4.x 范围(`>=4.57.0,<5`)。Transformers 5.x 可能使该- custom-code 模型生成无效的全零 codes。--### 调用 API--无参考音频生成:--```bash-curl -sS --fail-with-body \- -H 'Content-Type: application/json' \- -d '{- "model": "audio8/tts-0.6b",- "input": "你好,这是 Audio8 TTS 服务测试。",- "response_format": "wav",- "max_new_tokens": 256,- "temperature": 0.8,- "top_p": 0.95,- "top_k": 50- }' \- http://127.0.0.1:8010/v1/audio/speech \- -o audio8.wav-```--使用一个参考音色生成:+git clone https://github.com/Edge0-AI/Audio8_TTS.git+cd Audio8_TTS+python3 -m venv .venv+source .venv/bin/activate+pip install -r requirements.txt-```bash-curl -sS --fail-with-body \- -H 'Content-Type: application/json' \- -d '{- "model": "audio8/tts-0.6b",- "input": "这句话使用参考音色生成。",- "response_format": "wav",- "temperature": 0.8,- "top_p": 0.95,- "top_k": 50,- "references": [{- "audio_path": "/data/reference.wav",- "text": "参考录音对应的准确文本。"- }]- }' \- http://127.0.0.1:8010/v1/audio/speech \- -o audio8_clone.wav+python audio8_tts_infer.py \+ --model Edge0/Audio8-TTS-Realistic-0.6b \+ --text "你好,欢迎使用 Audio8 TTS。" \+ --output outputs/realistic.wav```-参考音频路径必须在服务运行环境中可见。当前适配器支持 TP=1 和每个请求一个 reference。-设置 `"stream": true` 可在生成过程中接收 SSE 音频分块。使用-`"response_format": "pcm"` 开销最低,每个事件的 `audio.data` 包含 Base64 编码音频。-流式输出默认每 12 个 codec frames 生成一块,使用 128 帧 decoder context 和 1 帧边界-guard。可通过 `AUDIO8_TTS_STREAM_CHUNK_FRAMES`、-`AUDIO8_TTS_STREAM_CONTEXT_FRAMES` 和 `AUDIO8_TTS_STREAM_GUARD_FRAMES` 调整。+音色克隆时追加 `--reference-audio reference.wav --reference-text "参考录音的准确文本。"`。-流式输出是服务端可选项:启动服务时设置 `AUDIO8_TTS_STREAM_ENABLED=1` 才会启用 SSE-流式;默认关闭,请求会在生成完成后一次性返回完整音频(请求中的 `"stream": true`-此时会被忽略)。默认非流式以获得更高的吞吐和稳定的显存占用。-`response_format` 为 `codes`/`codec`/`npy` 时始终非流式,直接返回 codec codes。+[安装、克隆与批量推理](guides/inference_zh.md) · [SFT 训练](guides/training_zh.md)-运行以下 smoke test 可以验证部署:+#### 0.6B 评测摘要-```bash-BASE_URL=http://127.0.0.1:8010 ./sglang_omni/scripts/smoke_test.sh-python3 ./sglang_omni/scripts/stream_smoke_test.py \- --base-url http://127.0.0.1:8010 \- --output /tmp/audio8_stream.wav-```+**Audio8-TTS-Preview** 与 **Audio8-TTS-Realistic**,均为 **0.6B**:-构建镜像时,请在安装 SGLang Omni 包及其 Python 依赖后追加-[`sglang_omni/Dockerfile.snippet`](sglang_omni/Dockerfile.snippet)。+| 基准 | 覆盖 | Preview 错误率 % ↓ | Realistic 错误率 % ↓ | Preview SIM ↑ | Realistic SIM ↑ |+|---|---|---:|---:|---:|---:|+| Seed-TTS | 英语 / 中文 | 1.23 | 1.25 | 0.682 | 0.704 |+| CV3 | 8 种共有语言 | 4.80 | 4.56 | — | 0.663 |+| CV3 | 9 种核心语言 | — | 4.86 | — | 0.661 |+| MiniMax | 24 种评测语言 | — | 5.89 | — | 0.754 |-## SFT 训练+错误率为百分数,越低越好;SIM 使用 0–1 范围,越高越好。两者不能直接衡量语音拟真感。-先安装训练依赖:+> **说明:** 数值为各语言的无权重宏平均。Preview 数据来自其[已发布 Model Card](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b#evaluation),SIM 由百分数换算为 0–1。两个版本分别评测;“—”表示 Preview 卡片未报告对应结果。Seed-TTS 平均英语 WER 与中文 CER,不含 hard Chinese;CV3 共有集为中、英、日、韩、德、法、西、意 8 种语言,9 语言集另含俄语,不含 hard 子集及 13 种扩展语言。MiniMax 平均全部 24 种评测语言,其中部分不在正式支持清单内。-```bash-pip install -r requirements-train.txt-```+[完整 Realistic 评测与比较说明](guides/models/realistic-0.6b_zh.md#评测) · [Preview 原始评测](guides/models/preview-0.6b_zh.md#评测结果)-### 1. 创建原始数据 manifest+</details>-目标 `audio` 必填;`reference_audio` 和 `reference_text` 可选,但必须同时出现。+<details>+<summary><strong>上一版本 · Preview 0.6B</strong></summary>-```json-{"id":"utt_001","text":"目标音频文本","audio":"audio/target.wav","reference_audio":"audio/reference.wav","reference_text":"参考音频文本"}-{"id":"utt_002","text":"另一条文本","audio":"audio/another.wav"}-```+Realistic 在 Preview 基础上继续改进。Preview 用户仍可使用原版模型、**11 种语言**的支持说明、原评测与固定试听页。-### 2. 预计算 codec indices+| 版本 | 模型 | 试听 | 版本指南 | 部署 |+|---|---|---|---|---|+| Preview 0.6B | [HF](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b) | [Demo](https://edge0-ai.github.io/Audio8_TTS/preview-0.6b/) | [说明](guides/models/preview-0.6b_zh.md) | [ONNX INT4](onnx_runtime/README_zh.md) |-```bash-python audio8_tts_prepare.py \- --input-jsonl data/train.jsonl \- --output-jsonl prepared_data/train.jsonl \- --batch-size 4-```+</details>-生成的 manifest 使用相对路径指向经过校验的 `[10, T]` NumPy 数组。除非传入-`--overwrite`,否则已有的有效数组会被复用。+</details>-### 3. 开始训练+----单卡:+<a id="family-01b"></a>-```bash-TRAIN_JSONL=prepared_data/train.jsonl \-NPROC_PER_NODE=1 \-bash audio8_tts_sft.sh-```--单机八卡:--```bash-TRAIN_JSONL=prepared_data/train.jsonl \-NPROC_PER_NODE=8 \-BATCH_SIZE=2 \-GRADIENT_ACCUMULATION_STEPS=8 \-bash audio8_tts_sft.sh-```+<details open>+<summary><strong>0.1B family</strong></summary>-多机训练时,在每个节点设置 `NNODES`、`NODE_RANK`、`MASTER_ADDR` 和 `MASTER_PORT`。常用超参-和输出路径可通过 `audio8_tts_sft.sh` 中列出的环境变量覆盖,也可以在命令末尾附加 Transformers-参数。--SFT 同时优化 slow semantic/EOS loss 和 fast codebook teacher-forcing loss。只训练其中一个分支时,-可设置 `FREEZE_SLOW_AR=true` 或 `FREEZE_FAST_AR=true`。导出目录仍可通过标准 `AutoModel` 和-`AutoProcessor` 接口配合 `trust_remote_code=True` 加载。--## 评测结果--Audio8 TTS Preview 仅有 **0.6B 参数**,是本次对比中规模最小的模型。尽管参数量仅为其他模型的-一小部分,它在以下基准测试中依然达到业界领先 TTS 模型的第一梯队水平。其中,Audio8 TTS-Preview 在 Seed-TTS 上取得了最低的英文 WER 和有竞争力的中文 CER,并在 CV3 多语言评测中-保持了具有竞争力的整体表现。+<div align="center">-WER/CER 越低越好,SIM 越高越好。Seed-TTS 的相似度统一显示为百分数。+<img src="assets/readme-01b-zh.svg" alt="Audio8-TTS-0.1B — 零样本音色克隆 TTS,还能做到多小?" width="960">-### Seed-TTS+[](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b)+[](https://huggingface.co/Edge0/audio8-TTS-0.1B-ONNX-INT8)+[](https://edge0-ai.github.io/Audio8_TTS/0.1B/)-| 模型 | 参数量 | EN WER / SIM | ZH CER / SIM | Hard ZH CER / SIM |-|---|---:|---:|---:|---:|-| **Audio8 TTS Preview** | **0.6B** | **1.506** / 63.2 | 0.950 / 73.1 | 11.510 / 68.7 |-| Fish S2 Pro | 4.6B | 1.607 / 64.6 | 1.038 / 73.8 | 10.149 / 70.1 |-| Higgs Audio v2 | 4.7B | 1.524 / 66.4 | **0.806** / 72.1 | 10.622 / 69.3 |-| CosyVoice3-1.5B | 1.5B | 2.22 / 72.0 | 1.12 / 78.1 | **5.83** / **75.8** |-| MOSS-TTS | 8.5B | 1.85 / 73.4 | 1.20 / 78.8 | - |-| VoxCPM2 | 2.3B | 1.84 / **75.3** | 0.97 / **79.5** | 8.13 / 75.3 |+</div>-+**Audio8-TTS-0.1B**:支持零样本音色克隆的小巧、便携语音合成模型。-### CV3 多语言错误率+面向更小的部署体积,**以中文和英语为主要支持语言**。专用 ONNX INT8 运行时支持 CPU 推理、流式输出及本地 Web/HTTP 服务。-| 模型 | 参数量 | zh | en | hard-zh | hard-en | ja | ko | de | es | fr | it | ru |-|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|-| **Audio8 TTS Preview** | **0.6B** | **3.205** | **3.128** | 10.535 | 5.997 | 7.205 | 4.223 | 3.447 | 3.641 | 8.790 | 4.790 | - |-| Fish S2 Pro | 4.6B | 3.600 | 3.493 | 10.588 | 7.349 | 5.139 | **4.111** | 3.605 | 2.972 | **8.600** | 4.229 | **4.702** |-| Higgs Audio v2 | 4.7B | 3.378 | 3.404 | 10.424 | **5.754** | **4.742** | 4.260 | **3.300** | **2.929** | 9.425 | **3.555** | 5.423 |-| CosyVoice3-1.5B | 1.5B | 3.91 | 4.99 | 9.77 | 10.55 | 7.57 | 5.69 | 6.43 | 4.47 | 11.8 | 10.5 | 6.64 |-| VoxCPM2 | 2.3B | 3.65 | 5.00 | **8.55** | 8.48 | 5.96 | 5.69 | 4.77 | 3.80 | 9.85 | 4.25 | 5.21 |+| 版本 | 定位与语言 | 模型 | 试听 | 版本指南 |+|---|---|---|---|---|+| Preview 0.1B | 小体积混合架构 · 以中英为主 | [HF](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b) | [Demo](https://edge0-ai.github.io/Audio8_TTS/0.1B/) | [说明](guides/models/preview-0.1b_zh.md) |-+[0.1B 快速开始](guides/models/preview-0.1b_zh.md#推理) · [CPU / ONNX 运行时](onnx_runtime_0_1b_int8/README_zh.md) · [0.1B 部署](guides/models/preview-0.1b_zh.md#部署) · [模型许可](guides/models/preview-0.1b_zh.md#许可)-参数量直接根据已发布的权重张量统计。MOSS-TTS 包含 8,489,841,664 个参数;VoxCPM2 主模型-包含 2,290,004,544 个参数,单独的 AudioVAE 不计入本次参数量对比。+</details>-Fish S2 Pro 因官方评测使用自身 normalizer 而重新评测。Higgs Audio v2 因未公布具体数值而自行-评测。其他基线均采用 [VoxCPM 官方仓库](https://github.com/OpenBMB/VoxCPM)汇总的各项目官方值。+----不同 normalizer 和评测器下的跨项目数值只能作为参考,不能视为严格同口径排名;评测覆盖也不-会将 Preview 的正式支持范围扩展到前述 11 种语言之外。+## 共用工具-## 限制与负责任使用+[安装与推理](guides/inference_zh.md) · [SFT 训练](guides/training_zh.md)-- 当前是 Preview 模型,多语言和中文方言覆盖仍然有限。-- 过长、噪声较大或转写不准确的参考音频可能降低稳定性和音色相似度。-- 合成语音可能被用于冒充或传播虚假信息。克隆他人声音前应获得许可,并在适当场景明确标注合成内容。-- 部署前应针对具体业务完成准确性、安全性和合规测试。+共用工具通过显式模型 ID 选择版本;请使用对应版本的快速开始和部署说明。-## 许可证与致谢+## SGLang Omni Serving-本仓库代码和模型权重采用 [Apache License 2.0](LICENSE),归属说明见 [NOTICE](NOTICE)。+SGLang Omni 提供 paged attention、动态 batching 和兼容 OpenAI 的语音 API。0.6B 系列与 0.1B 混合架构模型使用各自的配置;部署指南保留现有性能数据及对应型号、硬件与测试条件。-感谢 Fish Audio 团队公开 Fish S2 Pro 使用的 DualAR 架构。+[部署矩阵、安装与服务配置](guides/deployment_zh.md) · [API、流式输出及排错](guides/deployment_zh.md#sglang-omni-服务部署)-## Star 历史+## 许可与致谢-<a href="https://www.star-history.com/?type=date&repos=Audio8-AI%2FAudio8_TTS">- <picture>- <source media="(prefers-color-scheme: dark)" srcset="https://api.star-history.com/chart?repos=Audio8-AI/Audio8_TTS&type=date&theme=dark&legend=top-left&sealed_token=ShFu9kcwBvymYQ4SjQ_NhkplHrefNRbYVYCiBIvIxnaBLKbEQ1cjHQBs2kZm7K5LNMWpU13JxWgA6zpHvmwh49FokyJ26axmq-0gG8b68Q8IJCyUDZW1jQ" />- <source media="(prefers-color-scheme: light)" srcset="https://api.star-history.com/chart?repos=Audio8-AI/Audio8_TTS&type=date&legend=top-left&sealed_token=ShFu9kcwBvymYQ4SjQ_NhkplHrefNRbYVYCiBIvIxnaBLKbEQ1cjHQBs2kZm7K5LNMWpU13JxWgA6zpHvmwh49FokyJ26axmq-0gG8b68Q8IJCyUDZW1jQ" />- <img alt="Star History Chart" src="https://api.star-history.com/chart?repos=Audio8-AI/Audio8_TTS&type=date&legend=top-left&sealed_token=ShFu9kcwBvymYQ4SjQ_NhkplHrefNRbYVYCiBIvIxnaBLKbEQ1cjHQBs2kZm7K5LNMWpU13JxWgA6zpHvmwh49FokyJ26axmq-0gG8b68Q8IJCyUDZW1jQ" />- </picture>-</a>+仓库代码和两个 0.6B 版本采用 [Apache License 2.0](LICENSE)。0.1B 权重采用独立的[模型许可](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b#license),各模型及导出制品以自身附带的许可为准。归属说明见 [NOTICE](NOTICE)。感谢 Fish Audio 团队公开 Fish Audio S2 Pro 使用的 DualAR 架构。diff --git a/assets/badges/arxiv.svg b/assets/badges/arxiv.svgnew file mode 100644index 0000000..9120561--- /dev/null+++ b/assets/badges/arxiv.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="133" height="20" role="img" aria-label="arXiv: 2610.14776"><title>arXiv: 2610.14776</title><filter id="blur"><feGaussianBlur stdDeviation="16"/></filter><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="133" height="20" rx="3"/></clipPath><g clip-path="url(#r)"><rect width="56" height="20" fill="#f2f3f5"/><rect x="56" width="77" height="20" fill="#3ab2da"/><rect width="133" height="20" fill="url(#s)"/></g><g fill="#2A2B3D" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><image x="5" y="3" width="14" height="14" href="data:image/svg+xml;base64,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"/><g transform="scale(.1)"><text x="375" y="140" textLength="290" fill="#333">arXiv</text></g><g transform="scale(.1)"><text x="935" y="140" textLength="670">2610.14776</text></g></g></svg>diff --git a/assets/badges/demo-01b.svg b/assets/badges/demo-01b.svgnew file mode 100644index 0000000..896dcab--- /dev/null+++ b/assets/badges/demo-01b.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="147" height="20" role="img" aria-label="Demo: 0.1B Samples"><title>Demo: 0.1B Samples</title><filter id="blur"><feGaussianBlur stdDeviation="16"/></filter><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="147" height="20" rx="3"/></clipPath><g clip-path="url(#r)"><rect width="60" height="20" fill="#f2f3f5"/><rect x="60" width="87" height="20" fill="#6fdba1"/><rect width="147" height="20" fill="url(#s)"/></g><g fill="#2A2B3D" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><image x="5" y="3" width="14" height="14" href="data:image/svg+xml;base64,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"/><g transform="scale(.1)"><text x="395" y="140" textLength="330" fill="#333">Demo</text></g><g transform="scale(.1)"><text x="1025" y="140" textLength="770" fill="#333">0.1B Samples</text></g></g></svg>diff --git a/assets/badges/demo.svg b/assets/badges/demo.svgnew file mode 100644index 0000000..1a39bb1--- /dev/null+++ b/assets/badges/demo.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="151" height="20" role="img" aria-label="Demo: Audio Samples"><title>Demo: Audio Samples</title><filter id="blur"><feGaussianBlur stdDeviation="16"/></filter><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="151" height="20" rx="3"/></clipPath><g clip-path="url(#r)"><rect width="60" height="20" fill="#f2f3f5"/><rect x="60" width="91" height="20" fill="#6fdba1"/><rect width="151" height="20" fill="url(#s)"/></g><g fill="#2A2B3D" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><image x="5" y="3" width="14" height="14" href="data:image/svg+xml;base64,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"/><g transform="scale(.1)"><text x="395" y="140" textLength="330" fill="#333">Demo</text></g><g transform="scale(.1)"><text x="1045" y="140" textLength="810" fill="#333">Audio Samples</text></g></g></svg>diff --git a/assets/badges/github.svg b/assets/badges/github.svgnew file mode 100644index 0000000..493d095--- /dev/null+++ b/assets/badges/github.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="141" height="20" role="img" aria-label="GitHub: Audio8_TTS"><title>GitHub: Audio8_TTS</title><filter id="blur"><feGaussianBlur stdDeviation="16"/></filter><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="141" height="20" rx="3"/></clipPath><g clip-path="url(#r)"><rect width="64" height="20" fill="#f2f3f5"/><rect x="64" width="77" height="20" fill="#2a2b3d"/><rect width="141" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><image x="5" y="3" width="14" height="14" href="data:image/svg+xml;base64,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"/><g transform="scale(.1)"><text x="415" y="140" textLength="370" fill="#333">GitHub</text></g><g transform="scale(.1)"><text x="1015" y="140" textLength="670">Audio8_TTS</text></g></g></svg>diff --git a/assets/badges/huggingface-preview-01b.svg b/assets/badges/huggingface-preview-01b.svgnew file mode 100644index 0000000..ea57698--- /dev/null+++ b/assets/badges/huggingface-preview-01b.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="185" height="20" role="img" aria-label="Hugging Face: Preview 0.1B"><title>Hugging Face: Preview 0.1B</title><filter id="blur"><feGaussianBlur stdDeviation="16"/></filter><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="185" height="20" rx="3"/></clipPath><g clip-path="url(#r)"><rect width="102" height="20" fill="#f2f3f5"/><rect x="102" width="83" height="20" fill="#ff5a21"/><rect width="185" height="20" fill="url(#s)"/></g><g fill="#2A2B3D" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><image x="5" y="3" width="14" height="14" href="data:image/svg+xml;base64,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"/><g transform="scale(.1)"><text x="605" y="140" textLength="750" fill="#333">Hugging Face</text></g><g transform="scale(.1)"><text x="1425" y="140" textLength="730">Preview 0.1B</text></g></g></svg>diff --git a/assets/badges/huggingface-realistic-06b.svg b/assets/badges/huggingface-realistic-06b.svgnew file mode 100644index 0000000..d265985--- /dev/null+++ b/assets/badges/huggingface-realistic-06b.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="187" height="20" role="img" aria-label="Hugging Face: Realistic 0.6B"><title>Hugging Face: Realistic 0.6B</title><filter id="blur"><feGaussianBlur stdDeviation="16"/></filter><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="187" height="20" rx="3"/></clipPath><g clip-path="url(#r)"><rect width="102" height="20" fill="#f2f3f5"/><rect x="102" width="85" height="20" fill="#ff5a21"/><rect width="187" height="20" fill="url(#s)"/></g><g fill="#2A2B3D" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><image x="5" y="3" width="14" height="14" href="data:image/svg+xml;base64,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"/><g transform="scale(.1)"><text x="605" y="140" textLength="750" fill="#333">Hugging Face</text></g><g transform="scale(.1)"><text x="1435" y="140" textLength="750">Realistic 0.6B</text></g></g></svg>diff --git a/assets/badges/license.svg b/assets/badges/license.svgnew file mode 100644index 0000000..b71e7e5--- /dev/null+++ b/assets/badges/license.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="156" height="20" role="img" aria-label="Code License: Apache 2.0"><title>Code License: Apache 2.0</title><filter id="blur"><feGaussianBlur stdDeviation="16"/></filter><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="156" height="20" rx="3"/></clipPath><g clip-path="url(#r)"><rect width="83" height="20" fill="#f2f3f5"/><rect x="83" width="73" height="20" fill="#6fdba1"/><rect width="156" height="20" fill="url(#s)"/></g><g fill="#2A2B3D" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><g transform="scale(.1)"><text x="425" y="140" textLength="730" fill="#333">Code License</text></g><g transform="scale(.1)"><text x="1185" y="140" textLength="630" fill="#333">Apache 2.0</text></g></g></svg>diff --git a/assets/badges/onnx-01b.svg b/assets/badges/onnx-01b.svgnew file mode 100644index 0000000..74e49dd--- /dev/null+++ b/assets/badges/onnx-01b.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="127" height="20" role="img" aria-label="ONNX: 0.1B INT8"><title>ONNX: 0.1B INT8</title><filter id="blur"><feGaussianBlur stdDeviation="16"/></filter><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="127" height="20" rx="3"/></clipPath><g clip-path="url(#r)"><rect width="60" height="20" fill="#f2f3f5"/><rect x="60" width="67" height="20" fill="#3ab2da"/><rect width="127" height="20" fill="url(#s)"/></g><g fill="#2A2B3D" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><image x="5" y="3" width="14" height="14" href="data:image/svg+xml;base64,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"/><g transform="scale(.1)"><text x="395" y="140" textLength="330" fill="#333">ONNX</text></g><g transform="scale(.1)"><text x="925" y="140" textLength="570">0.1B INT8</text></g></g></svg>diff --git a/assets/readme-01b-zh.svg b/assets/readme-01b-zh.svgnew file mode 100644index 0000000..9d7fd98--- /dev/null+++ b/assets/readme-01b-zh.svg@@ -0,0 +1,10 @@+<svg xmlns="http://www.w3.org/2000/svg" width="960" height="294" viewBox="0 0 960 294" role="img" aria-labelledby="title desc">+ <title id="title">Audio8-TTS-0.1B</title>+ <desc id="desc">零样本音色克隆 TTS, 还能做到多小?</desc>+ <rect width="960" height="294" fill="#ffffff"/>+ <g text-anchor="middle" font-family="-apple-system, BlinkMacSystemFont, 'Segoe UI', Helvetica, Arial, 'PingFang SC', 'Microsoft YaHei', sans-serif">+ <text x="480" y="80" font-size="24" font-weight="500" fill="#6e6e73">Audio8-TTS-0.1B</text>+ <text x="480" y="148" font-size="50" font-weight="600" letter-spacing="-1.2" fill="#1d1d1f">零样本音色克隆 TTS,</text>+ <text x="480" y="206" font-size="50" font-weight="600" letter-spacing="-1.2" fill="#1d1d1f">还能做到多小?</text>+ </g>+</svg>diff --git a/assets/readme-01b.svg b/assets/readme-01b.svgnew file mode 100644index 0000000..8200bd6--- /dev/null+++ b/assets/readme-01b.svg@@ -0,0 +1,10 @@+<svg xmlns="http://www.w3.org/2000/svg" width="960" height="294" viewBox="0 0 960 294" role="img" aria-labelledby="title desc">+ <title id="title">Audio8-TTS-0.1B</title>+ <desc id="desc">How small a zero-shot cloning TTS can be?</desc>+ <rect width="960" height="294" fill="#ffffff"/>+ <g text-anchor="middle" font-family="-apple-system, BlinkMacSystemFont, 'Segoe UI', Helvetica, Arial, 'PingFang SC', 'Microsoft YaHei', sans-serif">+ <text x="480" y="80" font-size="24" font-weight="500" fill="#6e6e73">Audio8-TTS-0.1B</text>+ <text x="480" y="148" font-size="50" font-weight="600" letter-spacing="-1.2" fill="#1d1d1f">How small a zero-shot</text>+ <text x="480" y="206" font-size="50" font-weight="600" letter-spacing="-1.2" fill="#1d1d1f">cloning TTS can be?</text>+ </g>+</svg>diff --git a/assets/readme-06b-zh.svg b/assets/readme-06b-zh.svgnew file mode 100644index 0000000..6563b98--- /dev/null+++ b/assets/readme-06b-zh.svg@@ -0,0 +1,11 @@+<svg xmlns="http://www.w3.org/2000/svg" width="960" height="350" viewBox="0 0 960 350" role="img" aria-labelledby="title desc">+ <title id="title">Audio8-TTS-Realistic</title>+ <desc id="desc">更真实的语音。 更广的语言覆盖。 依然只有 0.6B。</desc>+ <rect width="960" height="350" fill="#ffffff"/>+ <g text-anchor="middle" font-family="-apple-system, BlinkMacSystemFont, 'Segoe UI', Helvetica, Arial, 'PingFang SC', 'Microsoft YaHei', sans-serif">+ <text x="480" y="80" font-size="24" font-weight="500" fill="#6e6e73">Audio8-TTS-Realistic</text>+ <text x="480" y="148" font-size="50" font-weight="600" letter-spacing="-1.2" fill="#1d1d1f">更真实的语音。</text>+ <text x="480" y="206" font-size="50" font-weight="600" letter-spacing="-1.2" fill="#1d1d1f">更广的语言覆盖。</text>+ <text x="480" y="264" font-size="50" font-weight="600" letter-spacing="-1.2" fill="#1d1d1f">依然只有 0.6B。</text>+ </g>+</svg>diff --git a/assets/readme-06b.svg b/assets/readme-06b.svgnew file mode 100644index 0000000..924eda7--- /dev/null+++ b/assets/readme-06b.svg@@ -0,0 +1,11 @@+<svg xmlns="http://www.w3.org/2000/svg" width="960" height="350" viewBox="0 0 960 350" role="img" aria-labelledby="title desc">+ <title id="title">Audio8-TTS-Realistic</title>+ <desc id="desc">More realistic speech. More languages. Still 0.6B.</desc>+ <rect width="960" height="350" fill="#ffffff"/>+ <g text-anchor="middle" font-family="-apple-system, BlinkMacSystemFont, 'Segoe UI', Helvetica, Arial, 'PingFang SC', 'Microsoft YaHei', sans-serif">+ <text x="480" y="80" font-size="24" font-weight="500" fill="#6e6e73">Audio8-TTS-Realistic</text>+ <text x="480" y="148" font-size="50" font-weight="600" letter-spacing="-1.2" fill="#1d1d1f">More realistic speech.</text>+ <text x="480" y="206" font-size="50" font-weight="600" letter-spacing="-1.2" fill="#1d1d1f">More languages.</text>+ <text x="480" y="264" font-size="50" font-weight="600" letter-spacing="-1.2" fill="#1d1d1f">Still 0.6B.</text>+ </g>+</svg>diff --git a/docs/0.1B/index.html b/docs/0.1B/index.htmlindex ec85e0d..84af05c 100644--- a/docs/0.1B/index.html+++ b/docs/0.1B/index.html@@ -8,13 +8,24 @@<meta property="og:title" content="Audio8-TTS-0.1B - Listening Preview"><meta property="og:description" content="Multilingual, Mandarin, and English zero-shot text-to-speech listening preview."><meta property="og:type" content="website">- <meta property="og:url" content="https://audio8-ai.github.io/Audio8_TTS/0.1B/">+ <meta property="og:url" content="https://edge0-ai.github.io/Audio8_TTS/0.1B/"><title>Audio8-TTS-0.1B - Listening Preview</title>- <link rel="canonical" href="https://audio8-ai.github.io/Audio8_TTS/0.1B/">+ <link rel="canonical" href="https://edge0-ai.github.io/Audio8_TTS/0.1B/"><link rel="icon" href="data:,"><link rel="stylesheet" href="../styles.css?v=20260819-multilingual-5">+ <link rel="stylesheet" href="../version-nav.css"></head><body class="flat-demo-page">+ <nav class="version-bar" aria-label="Audio8 model versions">+ <div class="shell version-inner">+ <a class="version-brand" href="../">Audio8 TTS <span>Listening room</span></a>+ <div class="version-switcher">+ <a href="../realistic-0.6b/">Realistic <span>0.6B</span></a>+ <a href="../preview-0.6b/">Preview <span>0.6B</span></a>+ <a href="../0.1B/" aria-current="page">Preview <span>0.1B</span></a>+ </div>+ </div>+ </nav><header class="masthead" id="top"><div class="shell project-header"><div class="title-row">@@ -22,7 +33,10 @@<h1>Audio8-TTS-0.1B</h1><p class="subtitle">Multilingual zero-shot text-to-speech listening preview.</p></div>- <a class="repo-link" href="https://github.com/Audio8-AI/Audio8_TTS" target="_blank" rel="noreferrer">GitHub Repository</a>+ <div class="version-resources">+ <a class="repo-link" href="https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b" target="_blank" rel="noreferrer">Model on Hugging Face</a>+ <a class="repo-link" href="https://github.com/Edge0-AI/Audio8_TTS" target="_blank" rel="noreferrer">GitHub Repository</a>+ </div></div><div class="overview-grid"><section aria-labelledby="overview-title">diff --git a/docs/index.html b/docs/index.htmlindex 90ecb84..dc4daf1 100644--- a/docs/index.html+++ b/docs/index.html@@ -1,64 +1,78 @@<!doctype html><html lang="en">- <head>- <meta charset="utf-8">- <meta name="viewport" content="width=device-width, initial-scale=1">- <meta name="color-scheme" content="light">- <meta name="description" content="Audio8-TTS-0.6B local research listening preview">- <title>Audio8-TTS-0.6B - Listening Preview</title>- <link rel="stylesheet" href="styles.css?v=20260728-small-codec">- </head>- <body>- <header class="masthead" id="top">- <div class="shell project-header">- <div class="title-row">- <div>- <h1>Audio8-TTS-0.6B</h1>- <p class="subtitle">Research listening preview for multilingual zero-shot text-to-speech.</p>- </div>- <a class="repo-link" href="https://github.com/Audio8-AI/Audio8_TTS" target="_blank" rel="noreferrer">GitHub Repository</a>- </div>- <div class="overview-grid">- <section aria-labelledby="overview-title">- <h2 id="overview-title">Overview</h2>- <p>- This page presents outputs from the current Audio8-TTS-0.6B preview checkpoint.- It includes the recommended language set, targeted pronunciation challenges,- tongue twisters, complete classical Chinese works, cross-lingual cloning, and- dense lexical tests.- </p>- </section>- <section aria-labelledby="references-title">- <h2 id="references-title">Reference audio</h2>- <p>- The listening set combines difficult in-the-wild speech with cleaner expressive- prompts. This mix separates robustness to imperfect recordings from pronunciation,- prosody, and cross-language speaker-identity transfer.- </p>- </section>- </div>- </div>- </header>+<head>+ <meta charset="utf-8">+ <meta name="viewport" content="width=device-width, initial-scale=1">+ <meta name="color-scheme" content="light">+ <meta name="theme-color" content="#ffffff">+ <title>Audio8-TTS-Realistic · 0.6B — Listening Demo</title>+ <meta name="description" content="Listen to Audio8-TTS-Realistic: expressive speech in a compact 0.6B model, with 23 supported languages. Explore 24 examples alongside Preview, Breeze-TTS-2, Qwen3-TTS and Higgs TTS 3.">+ <meta property="og:type" content="website">+ <meta property="og:title" content="Audio8-TTS-Realistic · 0.6B">+ <meta property="og:description" content="More realistic speech. More languages. Still 0.6B. Explore 24 voice-cloning examples.">+ <meta property="og:url" content="https://edge0-ai.github.io/Audio8_TTS/">+ <link rel="canonical" href="https://edge0-ai.github.io/Audio8_TTS/">+ <link rel="stylesheet" href="version-nav.css">+ <link rel="stylesheet" href="realistic-0.6b/styles.css?v=2">+</head>+<body>+ <a class="skip-link" href="#listening">Skip to listening examples</a>+ <header class="version-bar" id="top">+ <div class="shell version-inner">+ <a class="version-brand" href="./">Audio8 TTS</a>+ <nav class="version-switcher" aria-label="Model versions">+ <a href="realistic-0.6b/" aria-current="page">Realistic <span>0.6B</span></a>+ <a href="preview-0.6b/">Preview <span>0.6B</span></a>+ <a href="0.1B/">Preview <span>0.1B</span></a>+ </nav>+ </div>+ </header>- <nav class="section-nav" aria-label="Listening sections">- <div class="shell nav-inner">- <a href="#multilingual">Multilingual <span>11</span></a>- <a href="#pronunciation-challenges">Pronunciation <span>05</span></a>- <a href="#tongue-twisters">Tongue twisters <span>08</span></a>- <a href="#classical-chinese">Classical Chinese <span>04</span></a>- <a href="#cross-lingual">Cross-lingual <span>08</span></a>- <a href="#hard-cases">Hard cases <span>07</span></a>+ <main>+ <section class="hero shell" aria-labelledby="hero-title">+ <h1 id="hero-title">Audio8-TTS-Realistic</h1>+ <p class="hero-copy">More realistic speech.<br>More languages.<br>Still 0.6B.</p>+ <p class="hero-detail">Built on Audio8-TTS-Preview, with richer expression and support for <strong>23 languages</strong>.</p>+ <div class="hero-actions">+ <a class="button model-link" href="https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b"><img src="realistic-0.6b/icons/huggingface.svg" alt="" width="20" height="20">Hugging Face <span aria-hidden="true">↗</span></a>+ <a class="button github-link" href="https://github.com/Edge0-AI/Audio8_TTS"><img src="realistic-0.6b/icons/github.svg" alt="" width="18" height="18">GitHub <span aria-hidden="true">↗</span></a>+ <a class="button paper-link" href="https://arxiv.org/abs/2610.14776"><img src="realistic-0.6b/icons/arxiv.svg" alt="" width="18" height="18">Paper <span aria-hidden="true">↗</span></a>+ </div>+ <div class="hero-facts" aria-label="Model highlights">+ <span><strong>0.6B</strong> model parameters</span>+ <span><strong>23</strong> supported languages</span>+ <span><strong>12</strong> more than Preview</span></div>- </nav>+ </section>- <main id="demo-root">- <p class="loading shell" role="status">Loading the listening set...</p>- </main>+ <section class="listening shell" id="listening" aria-labelledby="listening-title">+ <div class="section-heading">+ <div><p class="eyebrow">Listening room</p><h2 id="listening-title">A closer listen.</h2></div>+ <span class="example-total">24 examples · 5 models</span>+ </div>+ <p class="section-copy">The same text and reference voice, rendered by five models. Start with the reference and explore the samples at your own pace.</p>+ <p class="listening-note">Preview is included to illustrate the improvements in Realistic. The other models offer a broader reference for exploring different aspects of speech. Each has its own strengths; these selected examples are not an overall ranking of the models.</p>+ <div class="listening-controls">+ <div class="filters" role="group" aria-label="Filter examples by language">+ <button type="button" data-language="all" aria-pressed="true">All <span>24</span></button>+ <button type="button" data-language="en" aria-pressed="false">English <span>16</span></button>+ <button type="button" data-language="zh" aria-pressed="false">Chinese <span>8</span></button>+ </div>+ <label class="jump-label">Jump to <select id="sample-jump" aria-label="Jump to an example"><option value="">an example</option></select></label>+ </div>+ <p id="result-count" class="result-count" role="status">Showing all 24 examples</p>+ <div id="demo-root"><p class="loading">Loading listening examples…</p></div>+ <noscript><p>Enable JavaScript to load the listening examples and audio players.</p></noscript>+ </section>- <footer class="site-footer">- <div class="shell"><p><strong>Audio8-TTS-0.6B</strong> / Local listening preview</p></div>- </footer>- <a class="back-to-top" href="#top" aria-label="Back to top" title="Back to top">↑</a>- <script src="app.js?v=20260728-small-codec-r7" defer></script>- </body>+ <aside class="notes shell" aria-labelledby="notes-title">+ <h2 id="notes-title">About these examples</h2>+ <p>These are curated English and Chinese voice-cloning examples. They illustrate speech realism and expression; they do not cover all 23 supported languages or represent a blind preference test. See the <a href="https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b#evaluation">model card</a> for language coverage and evaluation.</p>+ <p>Sizes use the projects’ published model-size conventions: Audio8 Realistic and Preview 0.6B, <a href="https://huggingface.co/BreezeBlue/Breeze-TTS-2">Breeze-TTS-2 ~3B</a>, <a href="https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base">Qwen3-TTS-12Hz-1.7B-Base 1.7B</a>, and <a href="https://huggingface.co/bosonai/higgs-tts-3-4b">Higgs TTS 3 ~4B</a>. Counting conventions differ; Audio8’s 0.6B excludes its separate audio codec.</p>+ </aside>+ </main>+ <footer class="site-footer shell"><span><strong>Audio8 TTS</strong> · Compact models, expressive speech.</span><a href="#top">Back to top ↑</a></footer>+ <script src="realistic-0.6b/data.js" defer></script>+ <script src="realistic-0.6b/app.js?v=2" defer></script>+</body></html>diff --git a/docs/preview-0.6b/app.js b/docs/preview-0.6b/app.jsnew file mode 100644index 0000000..ee5cd45--- /dev/null+++ b/docs/preview-0.6b/app.js@@ -0,0 +1,202 @@+const CASE_SECTIONS = [+ {+ id: "pronunciation-challenges",+ index: "02 / 06",+ title: "Pronunciation Challenges",+ description: "Short cases for irregular spelling, contextual readings, place names, and classical readings.",+ groups: [+ {+ title: "Selected challenges",+ match: (sample) => sample.group === "selected_short",+ order: [+ "classical_polyphone_zh_01",+ "cross_enref_zh_places_01",+ "classical_polyphone_zh_02",+ "chaos_tear_en_01",+ "chaos_schism_en_01",+ ],+ },+ ],+ },+ {+ id: "tongue-twisters",+ index: "03 / 06",+ title: "Tongue Twisters",+ description: "Dense consonants and repeated syllables in English and Mandarin.",+ groups: [+ { title: "English", match: (sample) => sample.group === "tongue_twisters_en" },+ { title: "Mandarin", match: (sample) => sample.group === "tongue_twisters_zh" },+ ],+ },+ {+ id: "classical-chinese",+ index: "04 / 06",+ title: "Complete Classical Chinese",+ description: "Complete works with sustained continuity, controlled phrasing, and full closing lines.",+ groups: [+ { title: "Full works", match: (sample) => sample.group === "classical_chinese" },+ ],+ },+ {+ id: "cross-lingual",+ index: "05 / 06",+ title: "Cross-Lingual Voice Cloning",+ description: "Reference and target languages cross in both directions while speaker identity remains the anchor.",+ groups: [+ { title: "Mandarin reference to English", match: (sample) => sample.direction === "zh_ref_to_en" },+ {+ title: "English reference to Mandarin",+ match: (sample) => sample.direction === "en_ref_to_zh" && sample.group !== "selected_short",+ },+ ],+ },+ {+ id: "hard-cases",+ index: "06 / 06",+ title: "Hard Cases",+ description: "Uncommon vocabulary, technical terminology, acronyms, and initialisms.",+ groups: [+ { title: "Difficult terms", match: (sample) => sample.hardCaseType === "difficult_terms_names" },+ { title: "Acronyms and initialisms", match: (sample) => sample.hardCaseType === "acronyms_initialisms" },+ ],+ },+];++const LANGUAGE_LABELS = {+ "zh-CN": "Mandarin",+ "en-US": "English",+ zh: "Mandarin",+ en: "English",+ ja: "Japanese",+};++function escapeHTML(value) {+ return String(value)+ .replaceAll("&", "&")+ .replaceAll("<", "<")+ .replaceAll(">", ">")+ .replaceAll('"', """)+ .replaceAll("'", "'");+}++function languageLabel(value) {+ return LANGUAGE_LABELS[value] || value;+}++function player(path, label) {+ return `<audio controls preload="none" playsinline aria-label="${escapeHTML(label)}"><source src="${escapeHTML(path)}" type="audio/mpeg"></audio>`;+}++function compareSamples(left, right, preferredOrder = []) {+ const leftPreferred = preferredOrder.indexOf(left.id);+ const rightPreferred = preferredOrder.indexOf(right.id);+ const leftRank = leftPreferred === -1 ? Number.MAX_SAFE_INTEGER : leftPreferred;+ const rightRank = rightPreferred === -1 ? Number.MAX_SAFE_INTEGER : rightPreferred;+ if (leftRank !== rightRank) return leftRank - rightRank;+ const languageRank = { "en-US": 0, "zh-CN": 1 };+ return (languageRank[left.targetLanguage] ?? 2) - (languageRank[right.targetLanguage] ?? 2) || left.order - right.order;+}++function multilingualRow(sample) {+ return `+ <article class="language-row" data-sample-id="${escapeHTML(sample.id)}">+ <header class="language-cell">+ <span class="language-index">${String(sample.order).padStart(2, "0")}</span>+ <div><h3>${escapeHTML(sample.language)}</h3><p>${escapeHTML(sample.nativeName)}</p></div>+ </header>+ <div class="language-target">+ <p class="field-label">Target text</p>+ <p lang="${escapeHTML(sample.targetLanguage)}">${escapeHTML(sample.targetText)}</p>+ </div>+ <section class="compact-audio" aria-label="Reference audio">+ <div class="compact-heading"><span>Reference</span><small>${escapeHTML(languageLabel(sample.reference.language))}</small></div>+ ${player(sample.reference.audio, `Reference audio for ${sample.language}`)}+ <p lang="${escapeHTML(sample.reference.language)}">${escapeHTML(sample.reference.text)}</p>+ </section>+ <section class="compact-audio output-audio" aria-label="Generated audio">+ <div class="compact-heading"><span>Audio8 output</span><small>${escapeHTML(sample.nativeName)}</small></div>+ ${player(sample.output.audio, `Audio8 output in ${sample.language}`)}+ </section>+ </article>`;+}++function renderMultilingual(samples) {+ return `+ <section class="demo-section multilingual-section" id="multilingual">+ <div class="shell">+ <header class="section-heading">+ <p class="section-index">01 / 06</p>+ <h2>Multilingual Capability</h2>+ <p>Zero-shot synthesis across the 11 languages recommended for the preview checkpoint.</p>+ </header>+ <div class="language-table-head" aria-hidden="true">+ <span>Language</span><span>Target text</span><span>Reference</span><span>Generated audio</span>+ </div>+ <div class="language-list">${samples.sort((a, b) => a.order - b.order).map(multilingualRow).join("")}</div>+ </div>+ </section>`;+}++function sampleCard(sample, number) {+ const referenceLanguage = languageLabel(sample.reference.language);+ const targetLanguage = languageLabel(sample.targetLanguage);+ const direction = sample.direction ? `${referenceLanguage} ref / ${targetLanguage} out` : targetLanguage;+ const expected = sample.expectedReading+ ? `<p class="expected"><strong>Expected reading</strong>${escapeHTML(sample.expectedReading)}</p>`+ : "";+ const source = sample.source+ ? `<p class="source"><strong>Source</strong><a href="${escapeHTML(sample.source.url)}" target="_blank" rel="noreferrer">${escapeHTML(sample.source.label)}</a></p>`+ : "";+ return `+ <article class="sample-card" id="${escapeHTML(sample.id)}" data-sample-id="${escapeHTML(sample.id)}">+ <header class="sample-head">+ <span class="sample-number">${String(number).padStart(2, "0")}</span>+ <div><h3>${escapeHTML(sample.title)}</h3><p>${escapeHTML(sample.listenFor)}</p></div>+ <span class="language-badge">${escapeHTML(direction)}</span>+ </header>+ <div class="target-block">+ <p class="field-label">Target text</p>+ <div><p class="target-copy" lang="${escapeHTML(sample.targetLanguage)}">${escapeHTML(sample.targetText)}</p>${expected}${source}</div>+ </div>+ <div class="audio-pair">+ <section class="audio-block reference-block" aria-label="Reference audio">+ <div class="audio-heading"><span>Reference audio</span><small>${escapeHTML(referenceLanguage)}</small></div>+ <div>${player(sample.reference.audio, `Reference audio for ${sample.title}`)}<p class="reference-copy" lang="${escapeHTML(sample.reference.language)}">${escapeHTML(sample.reference.text)}</p></div>+ </section>+ <section class="audio-block generated-block" aria-label="Audio8 generated audio">+ <div class="audio-heading"><span>Generated audio</span><small>${escapeHTML(targetLanguage)}</small></div>+ <div>${player(sample.output.audio, `Audio8 generated audio for ${sample.title}`)}<p class="model-name">Audio8-TTS-Preview-0.6B</p></div>+ </section>+ </div>+ </article>`;+}++function renderCaseSection(definition, samples, state) {+ const groups = definition.groups.map((group) => {+ const matching = samples.filter(group.match).sort((a, b) => compareSamples(a, b, group.order));+ return `<div class="subsection"><h3 class="subsection-title">${escapeHTML(group.title)} <span>${String(matching.length).padStart(2, "0")}</span></h3><div class="sample-list">${matching.map((sample) => sampleCard(sample, state.next++)).join("")}</div></div>`;+ }).join("");+ return `<section class="demo-section" id="${definition.id}"><div class="shell"><header class="section-heading"><p class="section-index">${definition.index}</p><h2>${escapeHTML(definition.title)}</h2><p>${escapeHTML(definition.description)}</p></header>${groups}</div></section>`;+}++function bindAudio() {+ const players = [...document.querySelectorAll("audio")];+ players.forEach((audio) => audio.addEventListener("play", () => {+ players.forEach((other) => { if (other !== audio && !other.paused) other.pause(); });+ }));+}++async function initialize() {+ const response = await fetch("data.json?v=20260728-small-codec-r7");+ if (!response.ok) throw new Error(`Unable to load preview data: ${response.status}`);+ const payload = await response.json();+ const state = { next: 1 };+ document.querySelector("#demo-root").innerHTML = renderMultilingual(payload.multilingual)+ + CASE_SECTIONS.map((section) => renderCaseSection(section, payload.cases, state)).join("");+ bindAudio();+}++initialize().catch((error) => {+ document.querySelector("#demo-root").innerHTML = `<p class="loading shell">${escapeHTML(error.message)}</p>`;+ console.error(error);+});diff --git a/docs/preview-0.6b/data.json b/docs/preview-0.6b/data.jsonnew file mode 100644index 0000000..571eee1--- /dev/null+++ b/docs/preview-0.6b/data.json@@ -0,0 +1,833 @@+{+ "cases": [+ {+ "order": 1,+ "id": "chaos_tear_en_01",+ "group": "selected_short",+ "direction": null,+ "hardCaseType": "english_irregular_pronunciation",+ "title": "Tear in Eye, Dress You'll Tear",+ "targetLanguage": "en-US",+ "targetText": "I will keep you, Susy, busy,\nMake your head with heat grow dizzy;\nTear in eye, your dress you'll tear;\nQueer, fair seer, hear my prayer.",+ "listenFor": "the contrasting pronunciations of tear in a complete stanza",+ "expectedReading": "tear (in the eye): /tɪr/; tear (a dress): /tɛr/.",+ "audioVersion": "20260724-chaos",+ "source": {+ "label": "Gerard Nolst Trenité, The Chaos (1922)",+ "url": "https://ncf.idallen.com/english.html"+ },+ "reference": {+ "language": "en",+ "text": "What these pictures? An elephant with uh so many legs, and what's the woman doing on this beach? I don't know.",+ "audio": "../audio/cases/reference/chaos_tear_en_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/chaos_tear_en_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 2,+ "id": "chaos_schism_en_01",+ "group": "selected_short",+ "direction": null,+ "hardCaseType": "english_irregular_pronunciation",+ "title": "Realm, Schedule, and Schism",+ "targetLanguage": "en-US",+ "targetText": "But mind trivial and vial,\nTripod, menial, denial,\nTroll and trolley, realm and ream,\nSchedule, mischief, schism, and scheme.",+ "listenFor": "vial, tripod, realm, schedule, and schism within a complete stanza",+ "expectedReading": "vial /ˈvaɪəl/; tripod /ˈtraɪpɑd/; realm /rɛlm/; schedule /ˈskɛdʒuːl/; schism /ˈskɪzəm/.",+ "audioVersion": "20260724-chaos",+ "source": {+ "label": "Gerard Nolst Trenité, The Chaos (1922)",+ "url": "https://ncf.idallen.com/english.html"+ },+ "reference": {+ "language": "en",+ "text": "Really needs more coffee or something, but I have no time to do any coffee because I have too many deadlines. I should pass everything tomorrow and I feel really very bad.",+ "audio": "../audio/cases/reference/chaos_schism_en_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/chaos_schism_en_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 5,+ "id": "classical_polyphone_zh_01",+ "group": "selected_short",+ "direction": null,+ "hardCaseType": "classical_readings",+ "title": "Frontier Song - Lu Lun",+ "targetLanguage": "zh-CN",+ "targetText": "月黑雁飞高,单于夜遁逃。欲将轻骑逐,大雪满弓刀。",+ "listenFor": "the readings of chanyu and qingqi within a complete quatrain",+ "expectedReading": "单于 (chányú), 轻骑 (qīngqí, modern standard).",+ "reference": {+ "language": "zh",+ "text": "根据最新气象资料,本周末北方多地将迎来降温,出行前请及时查看天气预报。",+ "audio": "../audio/cases/reference/classical_polyphone_zh_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/classical_polyphone_zh_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 7,+ "id": "cross_enref_zh_places_01",+ "group": "selected_short",+ "direction": "en_ref_to_zh",+ "hardCaseType": "place_names",+ "title": "Chinese Place Names in Context from an English Reference",+ "targetLanguage": "zh-CN",+ "targetText": "名单中包括安徽省的六安、蚌埠、亳州和歙县,以及浙江省的丽水。",+ "listenFor": "contextual Chinese place-name readings from an English reference",+ "expectedReading": "六安 (Lù'ān), 蚌埠 (Bèngbù), 亳州 (Bózhōu), 歙县 (Shèxiàn), 丽水 (Líshuǐ).",+ "audioVersion": "20260724-demo",+ "reference": {+ "language": "en",+ "text": "Someone must move that table. I keep walking around here and I've never had that problem. The one time I'm not paying attention to where I'm going, of course. The one time I'm not wearing shoes.",+ "audio": "../audio/cases/reference/cross_enref_zh_places_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/cross_enref_zh_places_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 8,+ "id": "classical_polyphone_zh_02",+ "group": "selected_short",+ "direction": null,+ "hardCaseType": "classical_readings",+ "title": "Song of the Frontier - Lu Lun",+ "targetLanguage": "zh-CN",+ "targetText": "林暗草惊风,将军夜引弓。平明寻白羽,没在石棱中。",+ "listenFor": "the classical reading of 没 and complete line delivery",+ "expectedReading": "没在石棱中: mò zài shí léng zhōng.",+ "reference": {+ "language": "zh",+ "text": "您别看这件事听着简单,真要做起来,还得把顺序一件一件理清楚。",+ "audio": "../audio/cases/reference/classical_polyphone_zh_02.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/classical_polyphone_zh_02.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 1,+ "id": "twister_zh_01",+ "group": "tongue_twisters_zh",+ "direction": null,+ "hardCaseType": null,+ "title": "Eight Hundred Standard-Bearers",+ "targetLanguage": "zh-CN",+ "targetText": "八百标兵奔北坡,炮兵并排北边跑;炮兵怕把标兵碰,标兵怕碰炮兵炮。",+ "listenFor": "b/p consonant boundaries and complete final syllables",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "至于什么山呢?神山世界中心啊,那四个字现在不让说了。标题上不能带。昨天带了以后直接拿下。所以说,就说就叫那那个山了,世界中心。",+ "audio": "../audio/cases/reference/twister_zh_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/twister_zh_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 2,+ "id": "twister_zh_02",+ "group": "tongue_twisters_zh",+ "direction": null,+ "hardCaseType": null,+ "title": "Niu Lang and Liu Niang",+ "targetLanguage": "zh-CN",+ "targetText": "牛郎恋刘娘,刘娘念牛郎;牛郎年年恋刘娘,刘娘年年念牛郎。",+ "listenFor": "n/l contrast and repeated name consistency",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "像我早上用完,今天一整天都是清清爽爽的。这男士洗面奶我也用过不少,大多洗完之后脸上特别干,紧绷得慌,还有一股很重的酒精味儿。",+ "audio": "../audio/cases/reference/twister_zh_02.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/twister_zh_02.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 3,+ "id": "twister_zh_03",+ "group": "tongue_twisters_zh",+ "direction": null,+ "hardCaseType": null,+ "title": "Red Carp, Green Carp, and Donkey",+ "targetLanguage": "zh-CN",+ "targetText": "红鲤鱼家有头小绿驴叫李屡屡,绿鲤鱼家有头小红驴叫吕里里。",+ "listenFor": "dense l initials and name separation",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "我的这位置能走吗?刘宗东一个Q,好角度Q了两个人。那这样老干爹在河道大获全胜。蝎子再一个一拳二重击配合着汤原的推。刘宗东点出一个赞,知道你换不了我,感觉就是我们输的那种情况呀。",+ "audio": "../audio/cases/reference/twister_zh_03.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/twister_zh_03.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 4,+ "id": "twister_zh_04",+ "group": "tongue_twisters_zh",+ "direction": null,+ "hardCaseType": null,+ "title": "Black Fertilizer",+ "targetLanguage": "zh-CN",+ "targetText": "黑化肥发灰,灰化肥发黑;黑化肥发灰会挥发,灰化肥挥发会发黑。",+ "listenFor": "h/f alternation without skipped words",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "首先,这个英雄啊,我确实没玩过。光看比赛的话,以为这英雄很难的,原来就是个大号得力。首先,他的被动:暴击额外造成百分之十的魔法伤害。",+ "audio": "../audio/cases/reference/twister_zh_04.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/twister_zh_04.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 5,+ "id": "twister_en_01",+ "group": "tongue_twisters_en",+ "direction": null,+ "hardCaseType": null,+ "title": "Peter Piper",+ "targetLanguage": "en-US",+ "targetText": "Peter Piper picked a peck of pickled peppers. If Peter Piper picked a peck of pickled peppers, where is the peck of pickled peppers Peter Piper picked?",+ "listenFor": "repeated p consonants, final clusters, and word completeness",+ "expectedReading": null,+ "reference": {+ "language": "en",+ "text": "You are really the the sunlight of my day. You, when I'm with you, I feel so much warmer and and better and safer and it's just the best feeling.",+ "audio": "../audio/cases/reference/twister_en_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/twister_en_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 6,+ "id": "twister_en_02",+ "group": "tongue_twisters_en",+ "direction": null,+ "hardCaseType": null,+ "title": "Seashells",+ "targetLanguage": "en-US",+ "targetText": "She sells seashells by the seashore. The shells she sells are surely seashells, so if she sells shells on the seashore, I'm sure she sells seashore shells.",+ "listenFor": "s/sh contrast and stable plural endings",+ "expectedReading": null,+ "reference": {+ "language": "en",+ "text": "Oh my gosh, I can't believe it! Are you serious? I can't believe I did that myself. What's going to be the next step from here?",+ "audio": "../audio/cases/reference/twister_en_02.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/twister_en_02.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 7,+ "id": "twister_en_03",+ "group": "tongue_twisters_en",+ "direction": null,+ "hardCaseType": null,+ "title": "Betty Botter",+ "targetLanguage": "en-US",+ "targetText": "Betty Botter bought some butter, but she said the butter's bitter. If I put it in my batter, it will make my batter bitter.",+ "listenFor": "rapid b/t alternation and distinct butter, bitter, and batter",+ "expectedReading": null,+ "reference": {+ "language": "en",+ "text": "Not an unusual sight, but she chose makeup so that'll be very smudgy for dad to deal with.",+ "audio": "../audio/cases/reference/twister_en_03.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/twister_en_03.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 8,+ "id": "twister_en_04",+ "group": "tongue_twisters_en",+ "direction": null,+ "hardCaseType": null,+ "title": "Sixth Sick Sheik",+ "targetLanguage": "en-US",+ "targetText": "The sixth sick sheik's sixth sheep's sick.",+ "listenFor": "sixth, sick, sheik, and sheep consonant clusters without deletion",+ "expectedReading": null,+ "reference": {+ "language": "en",+ "text": "What? Did you not see the stop sign? You just blew right through it. You could have T boned me and my wife right here. Are you kidding me? Dude, you need to watch where you're going.",+ "audio": "../audio/cases/reference/twister_en_04.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/twister_en_04.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 9,+ "id": "classical_zh_01",+ "group": "classical_chinese",+ "direction": null,+ "hardCaseType": null,+ "title": "Invitation to Wine - Li Bai",+ "targetLanguage": "zh-CN",+ "targetText": "君不见黄河之水天上来,奔流到海不复回。君不见高堂明镜悲白发,朝如青丝暮成雪。人生得意须尽欢,莫使金樽空对月。天生我材必有用,千金散尽还复来。烹羊宰牛且为乐,会须一饮三百杯。岑夫子,丹丘生,将进酒,杯莫停。与君歌一曲,请君为我倾耳听。钟鼓馔玉不足贵,但愿长醉不复醒。古来圣贤皆寂寞,惟有饮者留其名。陈王昔时宴平乐,斗酒十千恣欢谑。主人何为言少钱,径须沽取对君酌。五花马,千金裘,呼儿将出换美酒,与尔同销万古愁。",+ "listenFor": "complete-work continuity, sustained cadence, classical readings, and breath placement",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "这份报告明明昨天就该交,你现在才告诉我数据还没有核对完,怎么能这样?",+ "audio": "../audio/cases/reference/classical_zh_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/classical_zh_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 10,+ "id": "classical_zh_02",+ "group": "classical_chinese",+ "direction": null,+ "hardCaseType": null,+ "title": "Climbing High - Du Fu",+ "targetLanguage": "zh-CN",+ "targetText": "风急天高猿啸哀,渚清沙白鸟飞回。无边落木萧萧下,不尽长江滚滚来。万里悲秋常作客,百年多病独登台。艰难苦恨繁霜鬓,潦倒新停浊酒杯。",+ "listenFor": "complete regulated-verse structure, parallel phrasing, and final-line closure",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "我把那封信收进抽屉,很多话没有说出口,却一直留在心里。",+ "audio": "../audio/cases/reference/classical_zh_02.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/classical_zh_02.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 11,+ "id": "classical_zh_03",+ "group": "classical_chinese",+ "direction": null,+ "hardCaseType": null,+ "title": "Prelude to Water Melody - Su Shi",+ "targetLanguage": "zh-CN",+ "targetText": "明月几时有?把酒问青天。不知天上宫阙,今夕是何年。我欲乘风归去,又恐琼楼玉宇,高处不胜寒。起舞弄清影,何似在人间。转朱阁,低绮户,照无眠。不应有恨,何事长向别时圆?人有悲欢离合,月有阴晴圆缺,此事古难全。但愿人长久,千里共婵娟。",+ "listenFor": "complete-work phrasing, question cadence, and controlled transitions",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "清晨的薄雾沿着山谷慢慢散开,第一缕阳光落在安静的湖面上。",+ "audio": "../audio/cases/reference/classical_zh_03.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/classical_zh_03.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 12,+ "id": "classical_zh_04",+ "group": "classical_chinese",+ "direction": null,+ "hardCaseType": null,+ "title": "The Hard Road to Shu - Li Bai",+ "targetLanguage": "zh-CN",+ "targetText": "噫吁嚱,危乎高哉!蜀道之难,难于上青天!蚕丛及鱼凫,开国何茫然!尔来四万八千岁,不与秦塞通人烟。",+ "listenFor": "classical exclamations, uncommon names, and sustained opening cadence",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "我把那封信收进抽屉,很多话没有说出口,却一直留在心里。",+ "audio": "../audio/cases/reference/classical_zh_04.mp3?v=20260728-small-codec-r5"+ },+ "output": {+ "audio": "../audio/cases/output/classical_zh_04.mp3?v=20260728-small-codec-r5"+ }+ },+ {+ "order": 13,+ "id": "cross_zhref_en_01",+ "group": "cross_lingual",+ "direction": "zh_ref_to_en",+ "hardCaseType": null,+ "title": "Quiet Current",+ "targetLanguage": "en-US",+ "targetText": "A quiet current moved beneath the bridge as the first train crossed the river.",+ "listenFor": "English intelligibility with identity retained from a Mandarin reference",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "很多年以后,当我再次回到那座城市,才发现记忆里的街道和真实的风景已经完全不同。",+ "audio": "../audio/cases/reference/cross_zhref_en_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/cross_zhref_en_01.mp3?v=20260728-small-codec-r3"+ }+ },+ {+ "order": 14,+ "id": "cross_zhref_en_02",+ "group": "cross_lingual",+ "direction": "zh_ref_to_en",+ "hardCaseType": null,+ "title": "Station at Sunrise",+ "targetLanguage": "en-US",+ "targetText": "By sunrise, every window in the old station reflected a different shade of gold.",+ "listenFor": "natural English prosody from a Mandarin reference",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "我把蓝色的风筝带到公园,风一吹,它就飞得比屋顶还高。",+ "audio": "../audio/cases/reference/cross_zhref_en_02.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/cross_zhref_en_02.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 15,+ "id": "cross_zhref_en_03",+ "group": "cross_lingual",+ "direction": "zh_ref_to_en",+ "hardCaseType": null,+ "title": "Blue Folder",+ "targetLanguage": "en-US",+ "targetText": "Please place the blue folder beside the lamp before the afternoon meeting begins.",+ "listenFor": "stable consonants and speaker identity across languages",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "灯已经关了,大家都睡着了,我们轻一点说话,别把他们吵醒。",+ "audio": "../audio/cases/reference/cross_zhref_en_03.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/cross_zhref_en_03.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 16,+ "id": "cross_zhref_en_04",+ "group": "cross_lingual",+ "direction": "zh_ref_to_en",+ "hardCaseType": null,+ "title": "Museum Reopening",+ "targetLanguage": "en-US",+ "targetText": "The museum will reopen on Thursday after the final lighting system has been tested.",+ "listenFor": "English phrasing and timbre transfer from a Mandarin reference",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "前往杭州的旅客请注意,检票口已经开放,请带好随身物品有序进站。",+ "audio": "../audio/cases/reference/cross_zhref_en_04.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/cross_zhref_en_04.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 17,+ "id": "cross_enref_zh_01",+ "group": "cross_lingual",+ "direction": "en_ref_to_zh",+ "hardCaseType": null,+ "title": "Mountains After Rain",+ "targetLanguage": "zh-CN",+ "targetText": "雨停以后,远处的山峰慢慢从云层里显现出来。",+ "listenFor": "Mandarin intelligibility with identity retained from an English reference",+ "expectedReading": null,+ "reference": {+ "language": "en",+ "text": "The bakery opened before sunrise, and the first trays of bread were already cooling beside the window.",+ "audio": "../audio/cases/reference/cross_enref_zh_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/cross_enref_zh_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 18,+ "id": "cross_enref_zh_02",+ "group": "cross_lingual",+ "direction": "en_ref_to_zh",+ "hardCaseType": null,+ "title": "Meeting Folder",+ "targetLanguage": "zh-CN",+ "targetText": "请把桌上的蓝色文件夹交给下午来开会的同事。",+ "listenFor": "natural Mandarin prosody from an English reference",+ "expectedReading": null,+ "reference": {+ "language": "en",+ "text": "I left the spare keys beside the blue notebook, just in case you arrive before I get home.",+ "audio": "../audio/cases/reference/cross_enref_zh_02.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/cross_enref_zh_02.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 19,+ "id": "cross_enref_zh_03",+ "group": "cross_lingual",+ "direction": "en_ref_to_zh",+ "hardCaseType": null,+ "title": "First Train",+ "targetLanguage": "zh-CN",+ "targetText": "清晨第一班列车经过大桥时,河面上还笼罩着薄雾。",+ "listenFor": "stable Mandarin syllables and speaker identity across languages",+ "expectedReading": null,+ "reference": {+ "language": "en",+ "text": "After the storm passed, the children ran outside to inspect the branches scattered across the yard.",+ "audio": "../audio/cases/reference/cross_enref_zh_03.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/cross_enref_zh_03.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 20,+ "id": "cross_enref_zh_04",+ "group": "cross_lingual",+ "direction": "en_ref_to_zh",+ "hardCaseType": null,+ "title": "Museum Lighting",+ "targetLanguage": "zh-CN",+ "targetText": "博物馆将在照明系统完成测试后,于本周四重新开放。",+ "listenFor": "Mandarin phrasing and timbre transfer from an English reference",+ "expectedReading": null,+ "reference": {+ "language": "en",+ "text": "Our meeting has moved to Thursday afternoon, so please update the calendar before lunch.",+ "audio": "../audio/cases/reference/cross_enref_zh_04.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/cross_enref_zh_04.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 21,+ "id": "hard_terms_zh_01",+ "group": "hard_cases",+ "direction": null,+ "hardCaseType": "difficult_terms_names",+ "title": "Rare Chinese Characters",+ "targetLanguage": "zh-CN",+ "targetText": "耄耋老人谈起饕餮纹与狴犴像,言辞缱绻却不踟蹰,偶有龃龉也只是一笑置之。",+ "listenFor": "耄耋, 饕餮, 狴犴, 缱绻, 踟蹰, and 龃龉",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "我在窗边写下这段话,院子里的桂花刚刚开,空气里有一阵淡淡的香气。",+ "audio": "../audio/cases/reference/hard_terms_zh_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/hard_terms_zh_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 22,+ "id": "hard_terms_zh_02",+ "group": "hard_cases",+ "direction": null,+ "hardCaseType": "difficult_terms_names",+ "title": "Difficult Chinese Place Names",+ "targetLanguage": "zh-CN",+ "targetText": "从盱眙到郴州,再经亳州抵达台州,沿途地名中的生僻字最容易被误读。",+ "listenFor": "盱眙, 郴州, 亳州, and 台州",+ "expectedReading": null,+ "reference": {+ "language": "zh",+ "text": "欢迎来到海边小镇,沿着这条石板路往前走,就能看到今天开放的灯塔。",+ "audio": "../audio/cases/reference/hard_terms_zh_02.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/hard_terms_zh_02.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 23,+ "id": "hard_terms_en_01",+ "group": "hard_cases",+ "direction": null,+ "hardCaseType": "difficult_terms_names",+ "title": "Biomedical Vocabulary",+ "targetLanguage": "en-US",+ "targetText": "The otolaryngologist reviewed the electroencephalogram before discussing immunohistochemistry, pharmacokinetics, and the contraindications associated with hydroxychloroquine.",+ "listenFor": "long biomedical words without syllable deletion",+ "expectedReading": null,+ "reference": {+ "language": "en",+ "text": "Now I'm gonna relax, and just cuddle with a blanket and watch some anime. Ah, it was such a great day today. I'm just happy.",+ "audio": "../audio/cases/reference/hard_terms_en_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/hard_terms_en_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 25,+ "id": "hard_acronym_en_01",+ "group": "hard_cases",+ "direction": null,+ "hardCaseType": "acronyms_initialisms",+ "title": "Speech and AI Stack",+ "targetLanguage": "en-US",+ "targetText": "The ASR service sends a transcript to the LLM, while the TTS model renders the response on a GPU or an NPU.",+ "listenFor": "ASR, LLM, TTS, GPU, and NPU spoken letter by letter",+ "expectedReading": "A-S-R; L-L-M; T-T-S; G-P-U; N-P-U",+ "reference": {+ "language": "en",+ "text": "This is a disgusting donut that is filled with nasty beans, and I don't like beans. I don't like donuts either.",+ "audio": "../audio/cases/reference/hard_acronym_en_01.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/hard_acronym_en_01.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 26,+ "id": "hard_acronym_en_02",+ "group": "hard_cases",+ "direction": null,+ "hardCaseType": "acronyms_initialisms",+ "title": "Web Protocols",+ "targetLanguage": "en-US",+ "targetText": "The API returns JSON over HTTPS, stores the result in SQL, and exposes a URL for the HTML report.",+ "listenFor": "letterwise technical initialisms with natural phrasing",+ "expectedReading": "A-P-I; Jason; H-T-T-P-S; Sequel; U-R-L; H-T-M-L",+ "reference": {+ "language": "en",+ "text": "Oh my goodness, look at that! That is the sweetest, cutest thing ever. I love it.",+ "audio": "../audio/cases/reference/hard_acronym_en_02.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/hard_acronym_en_02.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 27,+ "id": "hard_acronym_en_03",+ "group": "hard_cases",+ "direction": null,+ "hardCaseType": "acronyms_initialisms",+ "title": "Acronyms Versus Initialisms",+ "targetLanguage": "en-US",+ "targetText": "NASA used the JWST and its NIRCam instrument to study a galaxy that the BBC later featured in a science report.",+ "listenFor": "NASA as a word and mixed initialism behavior",+ "expectedReading": "NASA word form; J-W-S-T; near-cam; B-B-C",+ "reference": {+ "language": "en",+ "text": "Your grades are good, you're good at basketball, you're good at football, I just I can't tell you what you do. I am so proud of you, I'm proud to call you my son.",+ "audio": "../audio/cases/reference/hard_acronym_en_03.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/hard_acronym_en_03.mp3?v=20260728-small-codec-r2"+ }+ },+ {+ "order": 28,+ "id": "hard_acronym_en_04",+ "group": "hard_cases",+ "direction": null,+ "hardCaseType": "acronyms_initialisms",+ "title": "Organizations and Roles",+ "targetLanguage": "en-US",+ "targetText": "The former CEO discussed NATO policy with CNN and the BBC before joining an IEEE panel in the UK.",+ "listenFor": "mixed acronym and initialism behavior in ordinary prose",+ "expectedReading": "C-E-O; NATO word form; C-N-N; B-B-C; I-triple-E; U-K",+ "reference": {+ "language": "en",+ "text": "I'm so sad and alone and scared and cold. Wonder if anybody's gonna take me home ever. I don't like it here at the pound.",+ "audio": "../audio/cases/reference/hard_acronym_en_04.mp3?v=20260728-small-codec-r2"+ },+ "output": {+ "audio": "../audio/cases/output/hard_acronym_en_04.mp3?v=20260728-small-codec-r2"+ }+ }+ ],+ "multilingual": [+ {+ "id": "multilingual_zh_01",+ "language": "Chinese",+ "nativeName": "中文",+ "targetLanguage": "zh-CN",+ "targetText": "大家好,语音合成是一件有趣的事情,我很高兴可以为你们展示多语言支持!",+ "order": 1,+ "reference": {+ "language": "zh",+ "text": "有人说啊,我们每天呢会和三百个陌生人擦肩而过。今天我想做一个特别的实验:当你收到来自陌生人的善意的时候,会有什么样的反应呢?",+ "audio": "../audio/multilingual/reference/multilingual_zh_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_zh_01.mp3?v=20260805-native-ref-r1"+ }+ },+ {+ "id": "multilingual_nl_01",+ "language": "Dutch",+ "nativeName": "Nederlands",+ "targetLanguage": "nl-NL",+ "targetText": "Hallo allemaal! Spraaksynthese is erg interessant, en ik laat jullie graag onze meertalige ondersteuning zien!",+ "order": 2,+ "reference": {+ "language": "nl",+ "text": "Goedemiddag, dit is het NOS Journaal in Makkelijke Taal en ik ben Jeroen Tjepkema. In dit journaal vertellen we dat veel mensen schade hebben aan hun nieren.",+ "audio": "../audio/multilingual/reference/multilingual_nl_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_nl_01.mp3?v=20260805-native-ref-r1"+ }+ },+ {+ "id": "multilingual_en_01",+ "language": "English",+ "nativeName": "English",+ "targetLanguage": "en-US",+ "targetText": "Hello everyone! Speech synthesis is fascinating, and I'm delighted to show you our multilingual support!",+ "order": 3,+ "reference": {+ "language": "en",+ "text": "The room is quiet now. Take a slow breath, close your eyes, and let the day fade gently away.",+ "audio": "../audio/multilingual/reference/multilingual_en_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_en_01.mp3?v=20260805-native-ref-r1"+ }+ },+ {+ "id": "multilingual_fr_01",+ "language": "French",+ "nativeName": "Français",+ "targetLanguage": "fr-FR",+ "targetText": "Bonjour à toutes et à tous ! La synthèse vocale est passionnante, et c'est avec grand plaisir que je vous présente notre prise en charge multilingue !",+ "order": 4,+ "reference": {+ "language": "fr",+ "text": "Françaises, Français, mes chers compatriotes, je m'adresse à vous ce soir en raison de la situation internationale et de ses conséquences pour notre pays et pour l'Europe.",+ "audio": "../audio/multilingual/reference/multilingual_fr_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_fr_01.mp3?v=20260805-native-ref-r1"+ }+ },+ {+ "id": "multilingual_de_01",+ "language": "German",+ "nativeName": "Deutsch",+ "targetLanguage": "de-DE",+ "targetText": "Hallo zusammen! Sprachsynthese ist eine spannende Sache, und ich freue mich sehr, euch unsere mehrsprachige Unterstützung präsentieren zu können!",+ "order": 5,+ "reference": {+ "language": "de",+ "text": "Willkommen zur Tagesschau in einfacher Sprache. Ich bin Susanne Stichler und das sind unsere Nachrichten.",+ "audio": "../audio/multilingual/reference/multilingual_de_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_de_01.mp3?v=20260805-native-ref-r1"+ }+ },+ {+ "id": "multilingual_it_01",+ "language": "Italian",+ "nativeName": "Italiano",+ "targetLanguage": "it-IT",+ "targetText": "Ciao a tutti! La sintesi vocale è davvero interessante e sono molto felice di potervi mostrare il nostro supporto multilingue!",+ "order": 6,+ "reference": {+ "language": "it",+ "text": "Ma poi quello che ci dicono è che i soldi non comprano la felicità e io credo che non sia vero. E anzi mi chiedo come fa a non esserci ancora chiarezza su un argomento così importante.",+ "audio": "../audio/multilingual/reference/multilingual_it_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_it_01.mp3?v=20260805-native-ref-r1"+ }+ },+ {+ "id": "multilingual_ja_01",+ "language": "Japanese",+ "nativeName": "日本語",+ "targetLanguage": "ja-JP",+ "targetText": "皆さん、こんにちは。音声合成は面白いものです。皆さんに多言語対応をご紹介できて、とても嬉しいです!",+ "order": 7,+ "reference": {+ "language": "ja",+ "text": "今日は少し寒いけれど、約束どおり駅前で会いましょう。温かい飲み物も忘れずに持ってきてね。",+ "audio": "../audio/multilingual/reference/multilingual_ja_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_ja_01.mp3?v=20260805-native-ref-r1"+ }+ },+ {+ "id": "multilingual_ko_01",+ "language": "Korean",+ "nativeName": "한국어",+ "targetLanguage": "ko-KR",+ "targetText": "안녕하세요, 여러분! 음성 합성은 정말 흥미로운 일이며, 여러분께 다국어 지원을 보여 드릴 수 있어 매우 기쁩니다!",+ "order": 8,+ "reference": {+ "language": "ko",+ "text": "이번 시간에는 첫 번째 주제인 한국의 상징에 대해 배워볼 건데요. 세계는 약 237개의 국가로 이루어져 있고",+ "audio": "../audio/multilingual/reference/multilingual_ko_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_ko_01.mp3?v=20260805-native-ref-r1"+ }+ },+ {+ "id": "multilingual_pl_01",+ "language": "Polish",+ "nativeName": "Polski",+ "targetLanguage": "pl-PL",+ "targetText": "Witam wszystkich! Synteza mowy to fascynujące zagadnienie i bardzo się cieszę, że mogę zaprezentować wam obsługę wielu języków!",+ "order": 9,+ "reference": {+ "language": "pl",+ "text": "Ale co by było, gdyby zamiast tego zadać banalne pytanie: drogi przyjacielu, a jakie kwiaty najbardziej lubi twoja żona?",+ "audio": "../audio/multilingual/reference/multilingual_pl_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_pl_01.mp3?v=20260805-native-ref-r1"+ }+ },+ {+ "id": "multilingual_es_01",+ "language": "Spanish",+ "nativeName": "Español",+ "targetLanguage": "es-ES",+ "targetText": "¡Hola a todos! La síntesis de voz es algo fascinante, y me alegra mucho poder mostrarles nuestra compatibilidad con varios idiomas.",+ "order": 10,+ "reference": {+ "language": "es",+ "text": "Bueno, yo pienso que eso sería bueno, ya que así tendremos más medicamentos, tendremos más posibilidades para los pacientes.",+ "audio": "../audio/multilingual/reference/multilingual_es_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_es_01.mp3?v=20260805-native-ref-r1"+ }+ },+ {+ "id": "multilingual_yue_01",+ "language": "Cantonese",+ "nativeName": "粤语",+ "targetLanguage": "yue-Hant-HK",+ "targetText": "大家好!語音合成係一件好有趣嘅事,我好高興可以為大家展示多語言支援!",+ "order": 11,+ "reference": {+ "language": "yue",+ "text": "李家超第四份施政報告主題係深化改革、心繫民生、發揮優勢、同創未來。綠色嘅封面寓意活力同埋政策延續,發展經濟同改善民生係今年施政報告嘅兩大主軸。",+ "audio": "../audio/multilingual/reference/multilingual_yue_01.mp3?v=20260805-native-ref-r1"+ },+ "output": {+ "audio": "../audio/multilingual/output/multilingual_yue_01.mp3?v=20260805-native-ref-r1"+ }+ }+ ]+}diff --git a/docs/preview-0.6b/index.html b/docs/preview-0.6b/index.htmlnew file mode 100644index 0000000..beacfee--- /dev/null+++ b/docs/preview-0.6b/index.html@@ -0,0 +1,84 @@+<!doctype html>+<html lang="en">+ <head>+ <meta charset="utf-8">+ <meta name="viewport" content="width=device-width, initial-scale=1">+ <meta name="color-scheme" content="light">+ <meta name="description" content="Audio8-TTS-Preview-0.6B: the original 11-language listening preview">+ <title>Audio8-TTS-Preview-0.6B - Listening Preview</title>+ <link rel="stylesheet" href="../styles.css?v=20260728-small-codec">+ <link rel="stylesheet" href="../version-nav.css">+ <link rel="canonical" href="https://edge0-ai.github.io/Audio8_TTS/preview-0.6b/">+ <meta property="og:title" content="Audio8-TTS-Preview-0.6B — Listening Preview">+ <meta property="og:description" content="The original Preview model, supporting 11 languages. Explore 43 listening examples with reference recordings.">+ <meta property="og:type" content="website">+ <meta property="og:url" content="https://edge0-ai.github.io/Audio8_TTS/preview-0.6b/">+ <link rel="icon" href="data:,">+ </head>+ <body>+ <nav class="version-bar" aria-label="Audio8 model versions">+ <div class="shell version-inner">+ <a class="version-brand" href="../">Audio8 TTS <span>Listening room</span></a>+ <div class="version-switcher">+ <a href="../realistic-0.6b/">Realistic <span>0.6B</span></a>+ <a href="../preview-0.6b/" aria-current="page">Preview <span>0.6B</span></a>+ <a href="../0.1B/">Preview <span>0.1B</span></a>+ </div>+ </div>+ </nav>+ <header class="masthead" id="top">+ <div class="shell project-header">+ <div class="title-row">+ <div>+ <h1>Audio8-TTS-Preview-0.6B</h1>+ <p class="subtitle">The original 0.6B Preview model · 11 supported languages · 43 listening examples.</p>+ </div>+ <div class="version-resources">+ <a class="repo-link" href="https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b" target="_blank" rel="noreferrer">Model on Hugging Face</a>+ <a class="repo-link" href="https://github.com/Edge0-AI/Audio8_TTS" target="_blank" rel="noreferrer">GitHub Repository</a>+ </div>+ </div>+ <div class="overview-grid">+ <section aria-labelledby="overview-title">+ <h2 id="overview-title">Overview</h2>+ <p>+ This page preserves outputs from Audio8-TTS-Preview-0.6B, supporting 11 languages.+ It includes the recommended language set, targeted pronunciation challenges,+ tongue twisters, complete classical Chinese works, cross-lingual cloning, and+ dense lexical tests.+ </p>+ </section>+ <section aria-labelledby="references-title">+ <h2 id="references-title">Reference audio</h2>+ <p>+ The listening set combines difficult in-the-wild speech with cleaner expressive+ prompts. This mix separates robustness to imperfect recordings from pronunciation,+ prosody, and cross-language speaker-identity transfer.+ </p>+ </section>+ </div>+ </div>+ </header>++ <nav class="section-nav" aria-label="Listening sections">+ <div class="shell nav-inner">+ <a href="#multilingual">Multilingual <span>11</span></a>+ <a href="#pronunciation-challenges">Pronunciation <span>05</span></a>+ <a href="#tongue-twisters">Tongue twisters <span>08</span></a>+ <a href="#classical-chinese">Classical Chinese <span>04</span></a>+ <a href="#cross-lingual">Cross-lingual <span>08</span></a>+ <a href="#hard-cases">Hard cases <span>07</span></a>+ </div>+ </nav>++ <main id="demo-root">+ <p class="loading shell" role="status">Loading the listening set...</p>+ </main>++ <footer class="site-footer">+ <div class="shell"><p><strong>Audio8-TTS-Preview-0.6B</strong> / Original 11-language listening preview</p></div>+ </footer>+ <a class="back-to-top" href="#top" aria-label="Back to top" title="Back to top">↑</a>+ <script src="app.js?v=20260728-small-codec-r7" defer></script>+ </body>+</html>diff --git a/docs/realistic-0.6b/app.js b/docs/realistic-0.6b/app.jsnew file mode 100644index 0000000..c24ca3c--- /dev/null+++ b/docs/realistic-0.6b/app.js@@ -0,0 +1,73 @@+/* Both the series landing page and the permanent demo use these same assets. */+(() => {+ "use strict";+ const assetBase = new URL(".", document.currentScript.src);+ const root = document.getElementById("demo-root");+ const escape = (text) => String(text ?? "").replace(/[&<>"']/g, c => ({"&":"&","<":"<",">":">",'"':""","'":"'"}[c]));+ const data = window.__DATA__;+ if (!data || !Array.isArray(data.samples) || !data.samples.length) {+ root.innerHTML = '<p class="audio-error" role="alert">The listening examples could not load. Please refresh the page.</p>';+ return;+ }+ const player = (sample, column) => {+ const model = data.models[column.key];+ const focus = column.key === "realistic";+ return `<section class="player${focus ? " focus" : ""}${column.key === "reference" ? " reference" : ""}" aria-label="${escape(model.label)} for ${escape(sample.label)}">+ <div class="player-heading"><h4>${escape(model.label)}</h4>${model.size ? `<span class="size">${escape(model.size)}</span>` : ""}${focus ? '<span class="focus-caption">Current release</span>' : ""}</div>+ <div><audio controls preload="none" playsinline aria-label="${escape(sample.label)} — ${escape(model.label)}" src="${escape(new URL(column.audio, assetBase).href)}"></audio><span class="duration">${Number(column.duration).toFixed(2)} s${column.key === "reference" ? " · Reference recording" : ""}</span></div>+ </section>`;+ };+ root.innerHTML = data.samples.map(sample => {+ const reference = {key:"reference",audio:sample.reference_audio,duration:sample.reference_duration};+ const columns = [reference, ...sample.columns.filter(c => c.key !== "realistic"), ...sample.columns.filter(c => c.key === "realistic")];+ return `<article class="sample-card" id="${escape(sample.id)}" data-language="${escape(sample.language)}">+ <div class="sample-head"><div class="sample-meta"><h3>${escape(sample.label)}</h3><span class="language-tag">${sample.language === "zh" ? "Chinese" : "English"}</span><a class="sample-link" href="#${escape(sample.id)}" aria-label="Link to ${escape(sample.label)}">#</a></div>+ <p class="prompt" lang="${escape(sample.language)}">${escape(sample.prompt)}</p>+ <details class="reference-text"><summary>Reference transcript</summary><p>${escape(sample.reference_text)}</p></details></div>+ <div class="players">${columns.map(column => player(sample,column)).join("")}</div>+ </article>`;+ }).join("");+ const audios = [...root.querySelectorAll("audio")];+ audios.forEach(audio => {+ audio.addEventListener("play", () => audios.forEach(other => { if (other !== audio && !other.paused) other.pause(); }));+ audio.addEventListener("error", () => {+ if (audio.parentElement.querySelector(".audio-error")) return;+ const message = document.createElement("p");+ message.className = "audio-error";+ message.textContent = "Audio could not load. Please refresh to try again.";+ message.setAttribute("role", "status");+ audio.after(message);+ });+ });+ const cards = [...root.querySelectorAll(".sample-card")];+ const filters = [...document.querySelectorAll(".filters button")];+ const jump = document.getElementById("sample-jump");+ const count = document.getElementById("result-count");+ function filter(language) {+ filters.forEach(button => button.setAttribute("aria-pressed", String(button.dataset.language === language)));+ cards.forEach(card => {+ card.hidden = language !== "all" && card.dataset.language !== language;+ if (card.hidden) card.querySelectorAll("audio").forEach(audio => audio.pause());+ });+ const visible = data.samples.filter(sample => language === "all" || sample.language === language);+ jump.innerHTML = '<option value="">an example</option>' + visible.map(sample => `<option value="${escape(sample.id)}">${escape(sample.label)} · ${sample.language === "zh" ? "Chinese" : "English"}</option>`).join("");+ count.textContent = language === "all" ? `Showing all ${visible.length} examples` : `Showing ${visible.length} ${language === "zh" ? "Chinese" : "English"} examples`;+ }+ filters.forEach(button => {+ const language = button.dataset.language;+ button.querySelector("span").textContent = data.samples.filter(sample => language === "all" || sample.language === language).length;+ button.addEventListener("click", () => filter(language));+ });+ jump.addEventListener("change", () => { if (jump.value) location.hash = jump.value; });+ const showHash = () => {+ const id = location.hash.slice(1);+ const target = cards.find(card => card.id === id);+ if (target) {+ if (target.hidden) filter("all");+ target.scrollIntoView();+ }+ };+ filter("all");+ window.addEventListener("hashchange", showHash);+ if (location.hash) requestAnimationFrame(showHash);+})();diff --git a/docs/realistic-0.6b/audio/sample-01-breeze.wav b/docs/realistic-0.6b/audio/sample-01-breeze.wavnew file mode 100644index 0000000..537c431Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-01-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-01-higgs.wav b/docs/realistic-0.6b/audio/sample-01-higgs.wavnew file mode 100644index 0000000..a342d67Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-01-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-01-preview.wav b/docs/realistic-0.6b/audio/sample-01-preview.wavnew file mode 100644index 0000000..9f6692aBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-01-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-01-qwen.wav b/docs/realistic-0.6b/audio/sample-01-qwen.wavnew file mode 100644index 0000000..3525798Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-01-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-01-realistic.wav b/docs/realistic-0.6b/audio/sample-01-realistic.wavnew file mode 100644index 0000000..f45abcdBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-01-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-01-reference.wav b/docs/realistic-0.6b/audio/sample-01-reference.wavnew file mode 100644index 0000000..4841be4Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-01-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-02-breeze.wav b/docs/realistic-0.6b/audio/sample-02-breeze.wavnew file mode 100644index 0000000..86c56b8Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-02-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-02-higgs.wav b/docs/realistic-0.6b/audio/sample-02-higgs.wavnew file mode 100644index 0000000..b5de0c5Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-02-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-02-preview.wav b/docs/realistic-0.6b/audio/sample-02-preview.wavnew file mode 100644index 0000000..f07b4d3Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-02-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-02-qwen.wav b/docs/realistic-0.6b/audio/sample-02-qwen.wavnew file mode 100644index 0000000..b810892Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-02-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-02-realistic.wav b/docs/realistic-0.6b/audio/sample-02-realistic.wavnew file mode 100644index 0000000..cd70ff2Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-02-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-02-reference.wav b/docs/realistic-0.6b/audio/sample-02-reference.wavnew file mode 100644index 0000000..91896c5Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-02-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-03-breeze.wav b/docs/realistic-0.6b/audio/sample-03-breeze.wavnew file mode 100644index 0000000..32403afBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-03-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-03-higgs.wav b/docs/realistic-0.6b/audio/sample-03-higgs.wavnew file mode 100644index 0000000..1770a97Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-03-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-03-preview.wav b/docs/realistic-0.6b/audio/sample-03-preview.wavnew file mode 100644index 0000000..c0cfe56Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-03-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-03-qwen.wav b/docs/realistic-0.6b/audio/sample-03-qwen.wavnew file mode 100644index 0000000..b57ba63Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-03-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-03-realistic.wav b/docs/realistic-0.6b/audio/sample-03-realistic.wavnew file mode 100644index 0000000..ef7b6b4Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-03-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-03-reference.wav b/docs/realistic-0.6b/audio/sample-03-reference.wavnew file mode 100644index 0000000..16ba13aBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-03-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-04-breeze.wav b/docs/realistic-0.6b/audio/sample-04-breeze.wavnew file mode 100644index 0000000..a13ad52Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-04-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-04-higgs.wav b/docs/realistic-0.6b/audio/sample-04-higgs.wavnew file mode 100644index 0000000..a477de0Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-04-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-04-preview.wav b/docs/realistic-0.6b/audio/sample-04-preview.wavnew file mode 100644index 0000000..7475bc3Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-04-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-04-qwen.wav b/docs/realistic-0.6b/audio/sample-04-qwen.wavnew file mode 100644index 0000000..4b73c94Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-04-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-04-realistic.wav b/docs/realistic-0.6b/audio/sample-04-realistic.wavnew file mode 100644index 0000000..5b2b934Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-04-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-04-reference.wav b/docs/realistic-0.6b/audio/sample-04-reference.wavnew file mode 100644index 0000000..661618dBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-04-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-05-breeze.wav b/docs/realistic-0.6b/audio/sample-05-breeze.wavnew file mode 100644index 0000000..4482477Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-05-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-05-higgs.wav b/docs/realistic-0.6b/audio/sample-05-higgs.wavnew file mode 100644index 0000000..f687a27Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-05-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-05-preview.wav b/docs/realistic-0.6b/audio/sample-05-preview.wavnew file mode 100644index 0000000..c6e9e17Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-05-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-05-qwen.wav b/docs/realistic-0.6b/audio/sample-05-qwen.wavnew file mode 100644index 0000000..c1412c7Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-05-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-05-realistic.wav b/docs/realistic-0.6b/audio/sample-05-realistic.wavnew file mode 100644index 0000000..bac1453Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-05-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-05-reference.wav b/docs/realistic-0.6b/audio/sample-05-reference.wavnew file mode 100644index 0000000..b4f6607Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-05-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-06-breeze.wav b/docs/realistic-0.6b/audio/sample-06-breeze.wavnew file mode 100644index 0000000..f3342ecBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-06-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-06-higgs.wav b/docs/realistic-0.6b/audio/sample-06-higgs.wavnew file mode 100644index 0000000..45f2fe4Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-06-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-06-preview.wav b/docs/realistic-0.6b/audio/sample-06-preview.wavnew file mode 100644index 0000000..4882a65Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-06-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-06-qwen.wav b/docs/realistic-0.6b/audio/sample-06-qwen.wavnew file mode 100644index 0000000..7910749Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-06-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-06-realistic.wav b/docs/realistic-0.6b/audio/sample-06-realistic.wavnew file mode 100644index 0000000..27909f1Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-06-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-06-reference.wav b/docs/realistic-0.6b/audio/sample-06-reference.wavnew file mode 100644index 0000000..340aa4dBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-06-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-07-breeze.wav b/docs/realistic-0.6b/audio/sample-07-breeze.wavnew file mode 100644index 0000000..232cfc0Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-07-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-07-higgs.wav b/docs/realistic-0.6b/audio/sample-07-higgs.wavnew file mode 100644index 0000000..e45b710Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-07-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-07-preview.wav b/docs/realistic-0.6b/audio/sample-07-preview.wavnew file mode 100644index 0000000..a0352fcBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-07-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-07-qwen.wav b/docs/realistic-0.6b/audio/sample-07-qwen.wavnew file mode 100644index 0000000..282f062Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-07-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-07-realistic.wav b/docs/realistic-0.6b/audio/sample-07-realistic.wavnew file mode 100644index 0000000..400ec5fBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-07-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-07-reference.wav b/docs/realistic-0.6b/audio/sample-07-reference.wavnew file mode 100644index 0000000..e8f1caeBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-07-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-08-breeze.wav b/docs/realistic-0.6b/audio/sample-08-breeze.wavnew file mode 100644index 0000000..ae13379Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-08-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-08-higgs.wav b/docs/realistic-0.6b/audio/sample-08-higgs.wavnew file mode 100644index 0000000..564318aBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-08-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-08-preview.wav b/docs/realistic-0.6b/audio/sample-08-preview.wavnew file mode 100644index 0000000..02c05f5Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-08-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-08-qwen.wav b/docs/realistic-0.6b/audio/sample-08-qwen.wavnew file mode 100644index 0000000..e0682c1Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-08-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-08-realistic.wav b/docs/realistic-0.6b/audio/sample-08-realistic.wavnew file mode 100644index 0000000..3c863a3Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-08-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-08-reference.wav b/docs/realistic-0.6b/audio/sample-08-reference.wavnew file mode 100644index 0000000..befb7eaBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-08-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-09-breeze.wav b/docs/realistic-0.6b/audio/sample-09-breeze.wavnew file mode 100644index 0000000..9bfac82Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-09-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-09-higgs.wav b/docs/realistic-0.6b/audio/sample-09-higgs.wavnew file mode 100644index 0000000..8b83793Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-09-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-09-preview.wav b/docs/realistic-0.6b/audio/sample-09-preview.wavnew file mode 100644index 0000000..95cb7d5Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-09-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-09-qwen.wav b/docs/realistic-0.6b/audio/sample-09-qwen.wavnew file mode 100644index 0000000..63cc7baBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-09-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-09-realistic.wav b/docs/realistic-0.6b/audio/sample-09-realistic.wavnew file mode 100644index 0000000..4a1f92fBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-09-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-09-reference.wav b/docs/realistic-0.6b/audio/sample-09-reference.wavnew file mode 100644index 0000000..a50a4a7Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-09-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-10-breeze.wav b/docs/realistic-0.6b/audio/sample-10-breeze.wavnew file mode 100644index 0000000..79d17d7Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-10-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-10-higgs.wav b/docs/realistic-0.6b/audio/sample-10-higgs.wavnew file mode 100644index 0000000..67e7961Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-10-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-10-preview.wav b/docs/realistic-0.6b/audio/sample-10-preview.wavnew file mode 100644index 0000000..b802369Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-10-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-10-qwen.wav b/docs/realistic-0.6b/audio/sample-10-qwen.wavnew file mode 100644index 0000000..1df51f5Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-10-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-10-realistic.wav b/docs/realistic-0.6b/audio/sample-10-realistic.wavnew file mode 100644index 0000000..367f098Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-10-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-10-reference.wav b/docs/realistic-0.6b/audio/sample-10-reference.wavnew file mode 100644index 0000000..1709053Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-10-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-11-breeze.wav b/docs/realistic-0.6b/audio/sample-11-breeze.wavnew file mode 100644index 0000000..bbbc059Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-11-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-11-higgs.wav b/docs/realistic-0.6b/audio/sample-11-higgs.wavnew file mode 100644index 0000000..640e1a1Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-11-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-11-preview.wav b/docs/realistic-0.6b/audio/sample-11-preview.wavnew file mode 100644index 0000000..a713964Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-11-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-11-qwen.wav b/docs/realistic-0.6b/audio/sample-11-qwen.wavnew file mode 100644index 0000000..20ad67dBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-11-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-11-realistic.wav b/docs/realistic-0.6b/audio/sample-11-realistic.wavnew file mode 100644index 0000000..6257e3aBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-11-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-11-reference.wav b/docs/realistic-0.6b/audio/sample-11-reference.wavnew file mode 100644index 0000000..8989a4aBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-11-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-12-breeze.wav b/docs/realistic-0.6b/audio/sample-12-breeze.wavnew file mode 100644index 0000000..2d518e9Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-12-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-12-higgs.wav b/docs/realistic-0.6b/audio/sample-12-higgs.wavnew file mode 100644index 0000000..618f553Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-12-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-12-preview.wav b/docs/realistic-0.6b/audio/sample-12-preview.wavnew file mode 100644index 0000000..949ee90Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-12-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-12-qwen.wav b/docs/realistic-0.6b/audio/sample-12-qwen.wavnew file mode 100644index 0000000..d5f0714Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-12-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-12-realistic.wav b/docs/realistic-0.6b/audio/sample-12-realistic.wavnew file mode 100644index 0000000..0d61bdaBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-12-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-12-reference.wav b/docs/realistic-0.6b/audio/sample-12-reference.wavnew file mode 100644index 0000000..12f0a0eBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-12-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-13-breeze.wav b/docs/realistic-0.6b/audio/sample-13-breeze.wavnew file mode 100644index 0000000..e16811dBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-13-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-13-higgs.wav b/docs/realistic-0.6b/audio/sample-13-higgs.wavnew file mode 100644index 0000000..c9935caBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-13-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-13-preview.wav b/docs/realistic-0.6b/audio/sample-13-preview.wavnew file mode 100644index 0000000..b714790Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-13-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-13-qwen.wav b/docs/realistic-0.6b/audio/sample-13-qwen.wavnew file mode 100644index 0000000..882e2d4Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-13-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-13-realistic.wav b/docs/realistic-0.6b/audio/sample-13-realistic.wavnew file mode 100644index 0000000..64ac875Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-13-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-13-reference.wav b/docs/realistic-0.6b/audio/sample-13-reference.wavnew file mode 100644index 0000000..8937046Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-13-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-14-breeze.wav b/docs/realistic-0.6b/audio/sample-14-breeze.wavnew file mode 100644index 0000000..7681e34Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-14-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-14-higgs.wav b/docs/realistic-0.6b/audio/sample-14-higgs.wavnew file mode 100644index 0000000..5565aa6Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-14-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-14-preview.wav b/docs/realistic-0.6b/audio/sample-14-preview.wavnew file mode 100644index 0000000..b7d1224Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-14-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-14-qwen.wav b/docs/realistic-0.6b/audio/sample-14-qwen.wavnew file mode 100644index 0000000..f4b14f3Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-14-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-14-realistic.wav b/docs/realistic-0.6b/audio/sample-14-realistic.wavnew file mode 100644index 0000000..10e48d8Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-14-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-14-reference.wav b/docs/realistic-0.6b/audio/sample-14-reference.wavnew file mode 100644index 0000000..90d136cBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-14-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-15-breeze.wav b/docs/realistic-0.6b/audio/sample-15-breeze.wavnew file mode 100644index 0000000..6ff7df0Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-15-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-15-higgs.wav b/docs/realistic-0.6b/audio/sample-15-higgs.wavnew file mode 100644index 0000000..961eaffBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-15-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-15-preview.wav b/docs/realistic-0.6b/audio/sample-15-preview.wavnew file mode 100644index 0000000..30f7920Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-15-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-15-qwen.wav b/docs/realistic-0.6b/audio/sample-15-qwen.wavnew file mode 100644index 0000000..9ce444aBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-15-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-15-realistic.wav b/docs/realistic-0.6b/audio/sample-15-realistic.wavnew file mode 100644index 0000000..410e942Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-15-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-15-reference.wav b/docs/realistic-0.6b/audio/sample-15-reference.wavnew file mode 100644index 0000000..904a48aBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-15-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-16-breeze.wav b/docs/realistic-0.6b/audio/sample-16-breeze.wavnew file mode 100644index 0000000..31c4de0Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-16-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-16-higgs.wav b/docs/realistic-0.6b/audio/sample-16-higgs.wavnew file mode 100644index 0000000..8b3b171Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-16-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-16-preview.wav b/docs/realistic-0.6b/audio/sample-16-preview.wavnew file mode 100644index 0000000..2841edcBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-16-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-16-qwen.wav b/docs/realistic-0.6b/audio/sample-16-qwen.wavnew file mode 100644index 0000000..7c1aca9Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-16-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-16-realistic.wav b/docs/realistic-0.6b/audio/sample-16-realistic.wavnew file mode 100644index 0000000..dd129c6Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-16-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-16-reference.wav b/docs/realistic-0.6b/audio/sample-16-reference.wavnew file mode 100644index 0000000..fb7fb6bBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-16-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-17-breeze.wav b/docs/realistic-0.6b/audio/sample-17-breeze.wavnew file mode 100644index 0000000..a7a2eb2Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-17-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-17-higgs.wav b/docs/realistic-0.6b/audio/sample-17-higgs.wavnew file mode 100644index 0000000..05b235fBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-17-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-17-preview.wav b/docs/realistic-0.6b/audio/sample-17-preview.wavnew file mode 100644index 0000000..9c71e67Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-17-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-17-qwen.wav b/docs/realistic-0.6b/audio/sample-17-qwen.wavnew file mode 100644index 0000000..7e7a181Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-17-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-17-realistic.wav b/docs/realistic-0.6b/audio/sample-17-realistic.wavnew file mode 100644index 0000000..d77e6d1Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-17-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-17-reference.wav b/docs/realistic-0.6b/audio/sample-17-reference.wavnew file mode 100644index 0000000..01c4dd6Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-17-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-18-breeze.wav b/docs/realistic-0.6b/audio/sample-18-breeze.wavnew file mode 100644index 0000000..a3e9869Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-18-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-18-higgs.wav b/docs/realistic-0.6b/audio/sample-18-higgs.wavnew file mode 100644index 0000000..1ffc434Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-18-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-18-preview.wav b/docs/realistic-0.6b/audio/sample-18-preview.wavnew file mode 100644index 0000000..e3c3229Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-18-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-18-qwen.wav b/docs/realistic-0.6b/audio/sample-18-qwen.wavnew file mode 100644index 0000000..a3b62b3Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-18-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-18-realistic.wav b/docs/realistic-0.6b/audio/sample-18-realistic.wavnew file mode 100644index 0000000..9c4f850Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-18-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-18-reference.wav b/docs/realistic-0.6b/audio/sample-18-reference.wavnew file mode 100644index 0000000..e1e7ed6Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-18-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-19-breeze.wav b/docs/realistic-0.6b/audio/sample-19-breeze.wavnew file mode 100644index 0000000..35ae403Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-19-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-19-higgs.wav b/docs/realistic-0.6b/audio/sample-19-higgs.wavnew file mode 100644index 0000000..020f52eBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-19-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-19-preview.wav b/docs/realistic-0.6b/audio/sample-19-preview.wavnew file mode 100644index 0000000..8963472Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-19-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-19-qwen.wav b/docs/realistic-0.6b/audio/sample-19-qwen.wavnew file mode 100644index 0000000..38a01e1Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-19-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-19-realistic.wav b/docs/realistic-0.6b/audio/sample-19-realistic.wavnew file mode 100644index 0000000..1fe753cBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-19-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-19-reference.wav b/docs/realistic-0.6b/audio/sample-19-reference.wavnew file mode 100644index 0000000..5103fb4Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-19-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-20-breeze.wav b/docs/realistic-0.6b/audio/sample-20-breeze.wavnew file mode 100644index 0000000..38a882bBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-20-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-20-higgs.wav b/docs/realistic-0.6b/audio/sample-20-higgs.wavnew file mode 100644index 0000000..6d300ebBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-20-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-20-preview.wav b/docs/realistic-0.6b/audio/sample-20-preview.wavnew file mode 100644index 0000000..8b16292Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-20-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-20-qwen.wav b/docs/realistic-0.6b/audio/sample-20-qwen.wavnew file mode 100644index 0000000..411db96Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-20-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-20-realistic.wav b/docs/realistic-0.6b/audio/sample-20-realistic.wavnew file mode 100644index 0000000..cf63b75Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-20-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-20-reference.wav b/docs/realistic-0.6b/audio/sample-20-reference.wavnew file mode 100644index 0000000..678ede9Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-20-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-21-breeze.wav b/docs/realistic-0.6b/audio/sample-21-breeze.wavnew file mode 100644index 0000000..f834edaBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-21-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-21-higgs.wav b/docs/realistic-0.6b/audio/sample-21-higgs.wavnew file mode 100644index 0000000..c3c17a0Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-21-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-21-preview.wav b/docs/realistic-0.6b/audio/sample-21-preview.wavnew file mode 100644index 0000000..22d7f01Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-21-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-21-qwen.wav b/docs/realistic-0.6b/audio/sample-21-qwen.wavnew file mode 100644index 0000000..c4bb3e8Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-21-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-21-realistic.wav b/docs/realistic-0.6b/audio/sample-21-realistic.wavnew file mode 100644index 0000000..9840171Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-21-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-21-reference.wav b/docs/realistic-0.6b/audio/sample-21-reference.wavnew file mode 100644index 0000000..519c314Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-21-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-22-breeze.wav b/docs/realistic-0.6b/audio/sample-22-breeze.wavnew file mode 100644index 0000000..6330b34Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-22-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-22-higgs.wav b/docs/realistic-0.6b/audio/sample-22-higgs.wavnew file mode 100644index 0000000..5a1a725Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-22-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-22-preview.wav b/docs/realistic-0.6b/audio/sample-22-preview.wavnew file mode 100644index 0000000..be7fe05Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-22-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-22-qwen.wav b/docs/realistic-0.6b/audio/sample-22-qwen.wavnew file mode 100644index 0000000..a033587Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-22-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-22-realistic.wav b/docs/realistic-0.6b/audio/sample-22-realistic.wavnew file mode 100644index 0000000..618ff8eBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-22-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-22-reference.wav b/docs/realistic-0.6b/audio/sample-22-reference.wavnew file mode 100644index 0000000..fbec795Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-22-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-23-breeze.wav b/docs/realistic-0.6b/audio/sample-23-breeze.wavnew file mode 100644index 0000000..a6317b0Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-23-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-23-higgs.wav b/docs/realistic-0.6b/audio/sample-23-higgs.wavnew file mode 100644index 0000000..c78458aBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-23-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-23-preview.wav b/docs/realistic-0.6b/audio/sample-23-preview.wavnew file mode 100644index 0000000..68c90c4Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-23-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-23-qwen.wav b/docs/realistic-0.6b/audio/sample-23-qwen.wavnew file mode 100644index 0000000..1f83abfBinary files /dev/null and b/docs/realistic-0.6b/audio/sample-23-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-23-realistic.wav b/docs/realistic-0.6b/audio/sample-23-realistic.wavnew file mode 100644index 0000000..dc78862Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-23-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-23-reference.wav b/docs/realistic-0.6b/audio/sample-23-reference.wavnew file mode 100644index 0000000..0bca703Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-23-reference.wav differdiff --git a/docs/realistic-0.6b/audio/sample-24-breeze.wav b/docs/realistic-0.6b/audio/sample-24-breeze.wavnew file mode 100644index 0000000..2290663Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-24-breeze.wav differdiff --git a/docs/realistic-0.6b/audio/sample-24-higgs.wav b/docs/realistic-0.6b/audio/sample-24-higgs.wavnew file mode 100644index 0000000..506ba58Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-24-higgs.wav differdiff --git a/docs/realistic-0.6b/audio/sample-24-preview.wav b/docs/realistic-0.6b/audio/sample-24-preview.wavnew file mode 100644index 0000000..8f98747Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-24-preview.wav differdiff --git a/docs/realistic-0.6b/audio/sample-24-qwen.wav b/docs/realistic-0.6b/audio/sample-24-qwen.wavnew file mode 100644index 0000000..e068828Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-24-qwen.wav differdiff --git a/docs/realistic-0.6b/audio/sample-24-realistic.wav b/docs/realistic-0.6b/audio/sample-24-realistic.wavnew file mode 100644index 0000000..95ed978Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-24-realistic.wav differdiff --git a/docs/realistic-0.6b/audio/sample-24-reference.wav b/docs/realistic-0.6b/audio/sample-24-reference.wavnew file mode 100644index 0000000..e1c4380Binary files /dev/null and b/docs/realistic-0.6b/audio/sample-24-reference.wav differdiff --git a/docs/realistic-0.6b/data.js b/docs/realistic-0.6b/data.jsnew file mode 100644index 0000000..acd990f--- /dev/null+++ b/docs/realistic-0.6b/data.js@@ -0,0 +1,895 @@+window.__DATA__ = {+ "title": "Audio8-TTS-Realistic",+ "models": {+ "reference": {+ "label": "Reference",+ "size": null+ },+ "preview": {+ "label": "Audio8-TTS-Preview",+ "size": "0.6B"+ },+ "realistic": {+ "label": "Audio8-TTS-Realistic",+ "size": "0.6B"+ },+ "breeze": {+ "label": "Breeze-TTS-2",+ "size": "~3B"+ },+ "qwen": {+ "label": "Qwen3-TTS",+ "size": "1.7B"+ },+ "higgs": {+ "label": "Higgs TTS 3",+ "size": "~4B"+ }+ },+ "samples": [+ {+ "id": "sample-01",+ "label": "Example 01",+ "language": "en",+ "prompt": "Listen up! You think this is a maybe? You need to be on that screen, and you need it now!",+ "reference_text": "Well, you better think again. You can't afford not to be on TV.",+ "reference_audio": "audio/sample-01-reference.wav",+ "reference_duration": 4.02,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-01-preview.wav",+ "duration": 6.83+ },+ {+ "key": "realistic",+ "audio": "audio/sample-01-realistic.wav",+ "duration": 7.15+ },+ {+ "key": "breeze",+ "audio": "audio/sample-01-breeze.wav",+ "duration": 4.24+ },+ {+ "key": "qwen",+ "audio": "audio/sample-01-qwen.wav",+ "duration": 5.44+ },+ {+ "key": "higgs",+ "audio": "audio/sample-01-higgs.wav",+ "duration": 5.6+ }+ ]+ },+ {+ "id": "sample-02",+ "label": "Example 02",+ "language": "en",+ "prompt": "Tomorrow, the whole city is going to know your name! So get up, get ready, and give me everything you've got!",+ "reference_text": "I can have you on the air tomorrow. You heard me right. Tomorrow. Better get ready to be famous, Albuquerque.",+ "reference_audio": "audio/sample-02-reference.wav",+ "reference_duration": 6.43,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-02-preview.wav",+ "duration": 6.69+ },+ {+ "key": "realistic",+ "audio": "audio/sample-02-realistic.wav",+ "duration": 7.62+ },+ {+ "key": "breeze",+ "audio": "audio/sample-02-breeze.wav",+ "duration": 6.56+ },+ {+ "key": "qwen",+ "audio": "audio/sample-02-qwen.wav",+ "duration": 6.4+ },+ {+ "key": "higgs",+ "audio": "audio/sample-02-higgs.wav",+ "duration": 7.68+ }+ ]+ },+ {+ "id": "sample-03",+ "label": "Example 03",+ "language": "en",+ "prompt": "You think I won't dig until I find the truth? I will tear this whole city apart, and when I do, you will answer for every lie.",+ "reference_text": "I just, I just couldn't prove it. He covered his tracks. He got that idiot at the copy shop to lie for him.",+ "reference_audio": "audio/sample-03-reference.wav",+ "reference_duration": 6.69,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-03-preview.wav",+ "duration": 7.52+ },+ {+ "key": "realistic",+ "audio": "audio/sample-03-realistic.wav",+ "duration": 9.24+ },+ {+ "key": "breeze",+ "audio": "audio/sample-03-breeze.wav",+ "duration": 7.52+ },+ {+ "key": "qwen",+ "audio": "audio/sample-03-qwen.wav",+ "duration": 7.12+ },+ {+ "key": "higgs",+ "audio": "audio/sample-03-higgs.wav",+ "duration": 7.8+ }+ ]+ },+ {+ "id": "sample-04",+ "label": "Example 04",+ "language": "en",+ "prompt": "I gave you a second chance when nobody else would. And this is what you do to me? You just signed the end of everything we built.",+ "reference_text": "And I saved him. And I shouldn't have. I took him into my own firm.",+ "reference_audio": "audio/sample-04-reference.wav",+ "reference_duration": 5.48,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-04-preview.wav",+ "duration": 8.08+ },+ {+ "key": "realistic",+ "audio": "audio/sample-04-realistic.wav",+ "duration": 7.76+ },+ {+ "key": "breeze",+ "audio": "audio/sample-04-breeze.wav",+ "duration": 5.84+ },+ {+ "key": "qwen",+ "audio": "audio/sample-04-qwen.wav",+ "duration": 5.52+ },+ {+ "key": "higgs",+ "audio": "audio/sample-04-higgs.wav",+ "duration": 6.96+ }+ ]+ },+ {+ "id": "sample-05",+ "label": "Example 05",+ "language": "en",+ "prompt": "Everyone keeps calling him a genius. But I know what he really is, and one day this whole town will know too.",+ "reference_text": "But not our Jimmy. Couldn't be precious Jimmy, stealing them blind, and he gets to be a lawyer. What a sick joke!",+ "reference_audio": "audio/sample-05-reference.wav",+ "reference_duration": 10.33,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-05-preview.wav",+ "duration": 8.87+ },+ {+ "key": "realistic",+ "audio": "audio/sample-05-realistic.wav",+ "duration": 9.2+ },+ {+ "key": "breeze",+ "audio": "audio/sample-05-breeze.wav",+ "duration": 7.44+ },+ {+ "key": "qwen",+ "audio": "audio/sample-05-qwen.wav",+ "duration": 6.48+ },+ {+ "key": "higgs",+ "audio": "audio/sample-05-higgs.wav",+ "duration": 7.36+ }+ ]+ },+ {+ "id": "sample-06",+ "label": "Example 06",+ "language": "en",+ "prompt": "Go ahead, stand there with that little smile. One day you will come to me because you need me, and I will remind you of this moment.",+ "reference_text": "You look down on me. You pity me. Walk away. That's right, Howard. You know why I didn't take the job? Cause it's too small.",+ "reference_audio": "audio/sample-06-reference.wav",+ "reference_duration": 7.24,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-06-preview.wav",+ "duration": 8.03+ },+ {+ "key": "realistic",+ "audio": "audio/sample-06-realistic.wav",+ "duration": 8.36+ },+ {+ "key": "breeze",+ "audio": "audio/sample-06-breeze.wav",+ "duration": 8.4+ },+ {+ "key": "qwen",+ "audio": "audio/sample-06-qwen.wav",+ "duration": 7.12+ },+ {+ "key": "higgs",+ "audio": "audio/sample-06-higgs.wav",+ "duration": 7.6+ }+ ]+ },+ {+ "id": "sample-07",+ "label": "Example 07",+ "language": "en",+ "prompt": "You think you understand what I am? You have not even seen the door open yet. When it does, everything you call reality will bend.",+ "reference_text": "I don't care about it. It's nothing to me. It's a bacterium. I travel in worlds you can't even imagine. You can't conceive of what I'm capable of.",+ "reference_audio": "audio/sample-07-reference.wav",+ "reference_duration": 10.82,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-07-preview.wav",+ "duration": 9.2+ },+ {+ "key": "realistic",+ "audio": "audio/sample-07-realistic.wav",+ "duration": 9.33+ },+ {+ "key": "breeze",+ "audio": "audio/sample-07-breeze.wav",+ "duration": 9.36+ },+ {+ "key": "qwen",+ "audio": "audio/sample-07-qwen.wav",+ "duration": 7.84+ },+ {+ "key": "higgs",+ "audio": "audio/sample-07-higgs.wav",+ "duration": 8.04+ }+ ]+ },+ {+ "id": "sample-08",+ "label": "Example 08",+ "language": "en",+ "prompt": "Oh, I am ready! You just wait—I'll go all night if I have to. This is going to be huge!",+ "reference_text": "I gotta hate. Yeah, I'ma hate. But I ain't have no problem with Dante doing his brokeback thing.",+ "reference_audio": "audio/sample-08-reference.wav",+ "reference_duration": 5.13,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-08-preview.wav",+ "duration": 5.94+ },+ {+ "key": "realistic",+ "audio": "audio/sample-08-realistic.wav",+ "duration": 7.2+ },+ {+ "key": "breeze",+ "audio": "audio/sample-08-breeze.wav",+ "duration": 5.2+ },+ {+ "key": "qwen",+ "audio": "audio/sample-08-qwen.wav",+ "duration": 4.24+ },+ {+ "key": "higgs",+ "audio": "audio/sample-08-higgs.wav",+ "duration": 5.12+ }+ ]+ },+ {+ "id": "sample-09",+ "label": "Example 09",+ "language": "en",+ "prompt": "This is incredible! I haven't tasted anything this good in years. You two are absolutely out of your minds—in the best way!",+ "reference_text": "Wow, so good! Don't suppose there's any tomato soup. Holy shit! You guys are something else.",+ "reference_audio": "audio/sample-09-reference.wav",+ "reference_duration": 7.48,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-09-preview.wav",+ "duration": 7.71+ },+ {+ "key": "realistic",+ "audio": "audio/sample-09-realistic.wav",+ "duration": 8.54+ },+ {+ "key": "breeze",+ "audio": "audio/sample-09-breeze.wav",+ "duration": 8.16+ },+ {+ "key": "qwen",+ "audio": "audio/sample-09-qwen.wav",+ "duration": 6.32+ },+ {+ "key": "higgs",+ "audio": "audio/sample-09-higgs.wav",+ "duration": 7.12+ }+ ]+ },+ {+ "id": "sample-10",+ "label": "Example 10",+ "language": "en",+ "prompt": "What an unbelievable twist! She finally figured it out, and now everything is about to spiral out of control.",+ "reference_text": "What a season! Amy figured out what I was up to, and now she's gonna use my DNA to make a multiverse waypoint.",+ "reference_audio": "audio/sample-10-reference.wav",+ "reference_duration": 8.13,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-10-preview.wav",+ "duration": 6.97+ },+ {+ "key": "realistic",+ "audio": "audio/sample-10-realistic.wav",+ "duration": 8.45+ },+ {+ "key": "breeze",+ "audio": "audio/sample-10-breeze.wav",+ "duration": 7.44+ },+ {+ "key": "qwen",+ "audio": "audio/sample-10-qwen.wav",+ "duration": 7.92+ },+ {+ "key": "higgs",+ "audio": "audio/sample-10-higgs.wav",+ "duration": 7.32+ }+ ]+ },+ {+ "id": "sample-11",+ "label": "Example 11",+ "language": "zh",+ "prompt": "我现在越来越不敢照镜子,总觉得镜子里那个人和以前完全不一样了。",+ "reference_text": "就是有点不相信自己,也不知道为什么会现在变成这样。我们以前是一个非常爱笑、非常热爱生活、非常爱锻炼的一个女孩子,就是我们。",+ "reference_audio": "audio/sample-11-reference.wav",+ "reference_duration": 11.17,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-11-preview.wav",+ "duration": 5.43+ },+ {+ "key": "realistic",+ "audio": "audio/sample-11-realistic.wav",+ "duration": 6.08+ },+ {+ "key": "breeze",+ "audio": "audio/sample-11-breeze.wav",+ "duration": 6.16+ },+ {+ "key": "qwen",+ "audio": "audio/sample-11-qwen.wav",+ "duration": 5.52+ },+ {+ "key": "higgs",+ "audio": "audio/sample-11-higgs.wav",+ "duration": 6.24+ }+ ]+ },+ {+ "id": "sample-12",+ "label": "Example 12",+ "language": "en",+ "prompt": "I have been left behind by everyone else, and the only thing that never walked away is this machine.",+ "reference_text": "My computer is the only thing that's been loyal to me and been there for me in my entire fucking life. So yes, I am serious.",+ "reference_audio": "audio/sample-12-reference.wav",+ "reference_duration": 10.66,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-12-preview.wav",+ "duration": 6.55+ },+ {+ "key": "realistic",+ "audio": "audio/sample-12-realistic.wav",+ "duration": 7.38+ },+ {+ "key": "breeze",+ "audio": "audio/sample-12-breeze.wav",+ "duration": 8.32+ },+ {+ "key": "qwen",+ "audio": "audio/sample-12-qwen.wav",+ "duration": 5.92+ },+ {+ "key": "higgs",+ "audio": "audio/sample-12-higgs.wav",+ "duration": 6.6+ }+ ]+ },+ {+ "id": "sample-13",+ "label": "Example 13",+ "language": "en",+ "prompt": "Every window is shut, every lock is checked twice, and still I don't feel safe inside this house.",+ "reference_text": "All the doors and windows are properly bolted. No one, not even me, went into the room last night. You have to believe me.",+ "reference_audio": "audio/sample-13-reference.wav",+ "reference_duration": 8.1,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-13-preview.wav",+ "duration": 6.46+ },+ {+ "key": "realistic",+ "audio": "audio/sample-13-realistic.wav",+ "duration": 7.06+ },+ {+ "key": "breeze",+ "audio": "audio/sample-13-breeze.wav",+ "duration": 6.88+ },+ {+ "key": "qwen",+ "audio": "audio/sample-13-qwen.wav",+ "duration": 6.48+ },+ {+ "key": "higgs",+ "audio": "audio/sample-13-higgs.wav",+ "duration": 7.68+ }+ ]+ },+ {+ "id": "sample-14",+ "label": "Example 14",+ "language": "zh",+ "prompt": "再大的风浪我也熬过来了,你现在看到的我,只会站得更直。",+ "reference_text": "我历经磨难,我依然站在你的面前,看着我,看着我。",+ "reference_audio": "audio/sample-14-reference.wav",+ "reference_duration": 9.43,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-14-preview.wav",+ "duration": 6.27+ },+ {+ "key": "realistic",+ "audio": "audio/sample-14-realistic.wav",+ "duration": 6.59+ },+ {+ "key": "breeze",+ "audio": "audio/sample-14-breeze.wav",+ "duration": 4.4+ },+ {+ "key": "qwen",+ "audio": "audio/sample-14-qwen.wav",+ "duration": 5.28+ },+ {+ "key": "higgs",+ "audio": "audio/sample-14-higgs.wav",+ "duration": 5.36+ }+ ]+ },+ {+ "id": "sample-15",+ "label": "Example 15",+ "language": "zh",+ "prompt": "你拿这个价去别处试试,看有没有人肯接你这摊子事,接完还得倒贴。",+ "reference_text": "咱平心而论啊,就你这个项目,搁正规医院咱得六千块钱,你在我这我给你两千块钱打九折,什么价格什么服务,你心里没点数吗?",+ "reference_audio": "audio/sample-15-reference.wav",+ "reference_duration": 9.47,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-15-preview.wav",+ "duration": 5.67+ },+ {+ "key": "realistic",+ "audio": "audio/sample-15-realistic.wav",+ "duration": 6.04+ },+ {+ "key": "breeze",+ "audio": "audio/sample-15-breeze.wav",+ "duration": 4.8+ },+ {+ "key": "qwen",+ "audio": "audio/sample-15-qwen.wav",+ "duration": 4.96+ },+ {+ "key": "higgs",+ "audio": "audio/sample-15-higgs.wav",+ "duration": 5.92+ }+ ]+ },+ {+ "id": "sample-16",+ "label": "Example 16",+ "language": "zh",+ "prompt": "我早就说清楚了,分工写得明明白白,你现在跟我装糊涂是不是?",+ "reference_text": "是不是有健忘症啊你?我里面跟你说没说?合伙人,人家出钱,我出技术,有毛病吗?",+ "reference_audio": "audio/sample-16-reference.wav",+ "reference_duration": 5.87,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-16-preview.wav",+ "duration": 4.23+ },+ {+ "key": "realistic",+ "audio": "audio/sample-16-realistic.wav",+ "duration": 4.92+ },+ {+ "key": "breeze",+ "audio": "audio/sample-16-breeze.wav",+ "duration": 5.12+ },+ {+ "key": "qwen",+ "audio": "audio/sample-16-qwen.wav",+ "duration": 5.12+ },+ {+ "key": "higgs",+ "audio": "audio/sample-16-higgs.wav",+ "duration": 4.72+ }+ ]+ },+ {+ "id": "sample-17",+ "label": "Example 17",+ "language": "zh",+ "prompt": "跟着我走,你才算真正回到家里,别再想那些外人了。",+ "reference_text": "跟你大爷在一起,咱们才是一家人。只有跟我在一起,永远在一起,你知道吗?",+ "reference_audio": "audio/sample-17-reference.wav",+ "reference_duration": 9.36,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-17-preview.wav",+ "duration": 6.59+ },+ {+ "key": "realistic",+ "audio": "audio/sample-17-realistic.wav",+ "duration": 6.41+ },+ {+ "key": "breeze",+ "audio": "audio/sample-17-breeze.wav",+ "duration": 3.84+ },+ {+ "key": "qwen",+ "audio": "audio/sample-17-qwen.wav",+ "duration": 4.24+ },+ {+ "key": "higgs",+ "audio": "audio/sample-17-higgs.wav",+ "duration": 5.72+ }+ ]+ },+ {+ "id": "sample-18",+ "label": "Example 18",+ "language": "zh",+ "prompt": "灯按下去还是没有光,屋里静得只听见自己的呼吸声。",+ "reference_text": "又顺着床沿摸到了台灯的开关,按了下去,屋子里仍然漆黑一片。",+ "reference_audio": "audio/sample-18-reference.wav",+ "reference_duration": 8.73,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-18-preview.wav",+ "duration": 5.62+ },+ {+ "key": "realistic",+ "audio": "audio/sample-18-realistic.wav",+ "duration": 7.43+ },+ {+ "key": "breeze",+ "audio": "audio/sample-18-breeze.wav",+ "duration": 6.48+ },+ {+ "key": "qwen",+ "audio": "audio/sample-18-qwen.wav",+ "duration": 5.68+ },+ {+ "key": "higgs",+ "audio": "audio/sample-18-higgs.wav",+ "duration": 7.36+ }+ ]+ },+ {+ "id": "sample-19",+ "label": "Example 19",+ "language": "zh",+ "prompt": "你看这个彬彬,才喝几罐就醉了,真的是太逊了!",+ "reference_text": "玩累了就直接睡觉,没问题的。",+ "reference_audio": "audio/sample-19-reference.wav",+ "reference_duration": 4.14,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-19-preview.wav",+ "duration": 4.92+ },+ {+ "key": "realistic",+ "audio": "audio/sample-19-realistic.wav",+ "duration": 4.64+ },+ {+ "key": "breeze",+ "audio": "audio/sample-19-breeze.wav",+ "duration": 5.12+ },+ {+ "key": "qwen",+ "audio": "audio/sample-19-qwen.wav",+ "duration": 4.72+ },+ {+ "key": "higgs",+ "audio": "audio/sample-19-higgs.wav",+ "duration": 4.8+ }+ ]+ },+ {+ "id": "sample-20",+ "label": "Example 20",+ "language": "en",+ "prompt": "Heroes come and go, but legends are forever.",+ "reference_text": "What can I say, Mamba out.",+ "reference_audio": "audio/sample-20-reference.wav",+ "reference_duration": 2.95,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-20-preview.wav",+ "duration": 2.79+ },+ {+ "key": "realistic",+ "audio": "audio/sample-20-realistic.wav",+ "duration": 3.67+ },+ {+ "key": "breeze",+ "audio": "audio/sample-20-breeze.wav",+ "duration": 4.0+ },+ {+ "key": "qwen",+ "audio": "audio/sample-20-qwen.wav",+ "duration": 3.68+ },+ {+ "key": "higgs",+ "audio": "audio/sample-20-higgs.wav",+ "duration": 3.28+ }+ ]+ },+ {+ "id": "sample-21",+ "label": "Example 21",+ "language": "zh",+ "prompt": "我在这行干了二十年,你觉得我会拿次品糊弄你?",+ "reference_text": "我开水果摊的,能卖给你生瓜蛋子?",+ "reference_audio": "audio/sample-21-reference.wav",+ "reference_duration": 3.64,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-21-preview.wav",+ "duration": 3.99+ },+ {+ "key": "realistic",+ "audio": "audio/sample-21-realistic.wav",+ "duration": 4.83+ },+ {+ "key": "breeze",+ "audio": "audio/sample-21-breeze.wav",+ "duration": 4.4+ },+ {+ "key": "qwen",+ "audio": "audio/sample-21-qwen.wav",+ "duration": 3.6+ },+ {+ "key": "higgs",+ "audio": "audio/sample-21-higgs.wav",+ "duration": 3.72+ }+ ]+ },+ {+ "id": "sample-22",+ "label": "Example 22",+ "language": "en",+ "prompt": "Every soul down there is a debt the world owes me. After tonight I answer to no one - not a master, not the dawn, not you. Surely you can see that.",+ "reference_text": "But if they die and I ascend, I won't have to rely on the parasite to walk in the sun. I'll be free, truly, completely free. Isn't that what you want?",+ "reference_audio": "audio/sample-22-reference.wav",+ "reference_duration": 18.11,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-22-preview.wav",+ "duration": 13.0+ },+ {+ "key": "realistic",+ "audio": "audio/sample-22-realistic.wav",+ "duration": 14.35+ },+ {+ "key": "breeze",+ "audio": "audio/sample-22-breeze.wav",+ "duration": 13.12+ },+ {+ "key": "qwen",+ "audio": "audio/sample-22-qwen.wav",+ "duration": 11.04+ },+ {+ "key": "higgs",+ "audio": "audio/sample-22-higgs.wav",+ "duration": 13.0+ }+ ]+ },+ {+ "id": "sample-23",+ "label": "Example 23",+ "language": "en",+ "prompt": "A truce, then. Until the brain is ash. Afterwards we can go back to hating each other properly.",+ "reference_text": "For now, perhaps we can put our differences aside, out of mutual self-interest.",+ "reference_audio": "audio/sample-23-reference.wav",+ "reference_duration": 7.01,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-23-preview.wav",+ "duration": 7.57+ },+ {+ "key": "realistic",+ "audio": "audio/sample-23-realistic.wav",+ "duration": 8.87+ },+ {+ "key": "breeze",+ "audio": "audio/sample-23-breeze.wav",+ "duration": 7.12+ },+ {+ "key": "qwen",+ "audio": "audio/sample-23-qwen.wav",+ "duration": 7.2+ },+ {+ "key": "higgs",+ "audio": "audio/sample-23-higgs.wav",+ "duration": 6.96+ }+ ]+ },+ {+ "id": "sample-24",+ "label": "Example 24",+ "language": "en",+ "prompt": "I know it's been a hard season. But every one of you chose this family, and not one of you rode away. That is what faith looks like.",+ "reference_text": "A moment. I just want to say thank you to Miss Karen, Bill, Lanny, Arthur. They are doing their jobs.",+ "reference_audio": "audio/sample-24-reference.wav",+ "reference_duration": 11.17,+ "columns": [+ {+ "key": "preview",+ "audio": "audio/sample-24-preview.wav",+ "duration": 8.41+ },+ {+ "key": "realistic",+ "audio": "audio/sample-24-realistic.wav",+ "duration": 10.45+ },+ {+ "key": "breeze",+ "audio": "audio/sample-24-breeze.wav",+ "duration": 9.28+ },+ {+ "key": "qwen",+ "audio": "audio/sample-24-qwen.wav",+ "duration": 8.08+ },+ {+ "key": "higgs",+ "audio": "audio/sample-24-higgs.wav",+ "duration": 9.04+ }+ ]+ }+ ]+};diff --git a/docs/realistic-0.6b/icons/arxiv.svg b/docs/realistic-0.6b/icons/arxiv.svgnew file mode 100644index 0000000..50cffbc--- /dev/null+++ b/docs/realistic-0.6b/icons/arxiv.svg@@ -0,0 +1 @@+<svg fill="#b31b1b" role="img" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><title>arXiv</title><path d="M3.8423 0a1.0037 1.0037 0 0 0-.922.6078c-.1536.3687-.0438.6275.2938 1.1113l6.9185 8.3597-1.0223 1.1058a1.0393 1.0393 0 0 0 .003 1.4229l1.2292 1.3135-5.4391 6.4444c-.2803.299-.4538.823-.2971 1.1986a1.0253 1.0253 0 0 0 .9585.635.9133.9133 0 0 0 .6891-.3405l5.783-6.126 7.4902 8.0051a.8527.8527 0 0 0 .6835.2597.9575.9575 0 0 0 .8777-.6138c.1577-.377-.017-.7502-.306-1.1407l-7.0518-8.3418 1.0632-1.13a.9626.9626 0 0 0 .0089-1.3165L4.6336.4639s-.3733-.4535-.768-.463zm0 .272h.0166c.2179.0052.4874.2715.5644.3639l.005.006.0052.0055 10.169 10.9905a.6915.6915 0 0 1-.0072.945l-1.0666 1.133-1.4982-1.7724-8.5994-10.39c-.3286-.472-.352-.6183-.2592-.841a.7307.7307 0 0 1 .6704-.4401Zm14.341 1.5701a.877.877 0 0 0-.6554.2418l-5.6962 6.1584 1.6944 1.8319 5.3089-6.5138c.3251-.4335.479-.6603.3247-1.0292a1.1205 1.1205 0 0 0-.9763-.689zm-7.6557 12.2823 1.3186 1.4135-5.7864 6.1295a.6494.6494 0 0 1-.4959.26.7516.7516 0 0 1-.706-.4669c-.1119-.2682.0359-.6864.2442-.9083l.0051-.0055.0047-.0055z"/></svg>\ No newline at end of filediff --git a/docs/realistic-0.6b/icons/github.svg b/docs/realistic-0.6b/icons/github.svgnew file mode 100644index 0000000..81920ca--- /dev/null+++ b/docs/realistic-0.6b/icons/github.svg@@ -0,0 +1 @@+<svg fill="#181717" role="img" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><title>GitHub</title><path d="M12 .297c-6.63 0-12 5.373-12 12 0 5.303 3.438 9.8 8.205 11.385.6.113.82-.258.82-.577 0-.285-.01-1.04-.015-2.04-3.338.724-4.042-1.61-4.042-1.61C4.422 18.07 3.633 17.7 3.633 17.7c-1.087-.744.084-.729.084-.729 1.205.084 1.838 1.236 1.838 1.236 1.07 1.835 2.809 1.305 3.495.998.108-.776.417-1.305.76-1.605-2.665-.3-5.466-1.332-5.466-5.93 0-1.31.465-2.38 1.235-3.22-.135-.303-.54-1.523.105-3.176 0 0 1.005-.322 3.3 1.23.96-.267 1.98-.399 3-.405 1.02.006 2.04.138 3 .405 2.28-1.552 3.285-1.23 3.285-1.23.645 1.653.24 2.873.12 3.176.765.84 1.23 1.91 1.23 3.22 0 4.61-2.805 5.625-5.475 5.92.42.36.81 1.096.81 2.22 0 1.606-.015 2.896-.015 3.286 0 .315.21.69.825.57C20.565 22.092 24 17.592 24 12.297c0-6.627-5.373-12-12-12"/></svg>\ No newline at end of filediff --git a/docs/realistic-0.6b/icons/huggingface.svg b/docs/realistic-0.6b/icons/huggingface.svgnew file mode 100644index 0000000..43c5d3c--- /dev/null+++ b/docs/realistic-0.6b/icons/huggingface.svg@@ -0,0 +1,37 @@+<svg xmlns="http://www.w3.org/2000/svg" width="95" height="88" fill="none">+ <path fill="#FFD21E" d="M47.21 76.5a34.75 34.75 0 1 0 0-69.5 34.75 34.75 0 0 0 0 69.5Z" />+ <path+ fill="#FF9D0B"+ d="M81.96 41.75a34.75 34.75 0 1 0-69.5 0 34.75 34.75 0 0 0 69.5 0Zm-73.5 0a38.75 38.75 0 1 1 77.5 0 38.75 38.75 0 0 1-77.5 0Z"+ />+ <path+ fill="#3A3B45"+ d="M58.5 32.3c1.28.44 1.78 3.06 3.07 2.38a5 5 0 1 0-6.76-2.07c.61 1.15 2.55-.72 3.7-.32ZM34.95 32.3c-1.28.44-1.79 3.06-3.07 2.38a5 5 0 1 1 6.76-2.07c-.61 1.15-2.56-.72-3.7-.32Z"+ />+ <path+ fill="#FF323D"+ d="M46.96 56.29c9.83 0 13-8.76 13-13.26 0-2.34-1.57-1.6-4.09-.36-2.33 1.15-5.46 2.74-8.9 2.74-7.19 0-13-6.88-13-2.38s3.16 13.26 13 13.26Z"+ />+ <path+ fill="#3A3B45"+ fill-rule="evenodd"+ d="M39.43 54a8.7 8.7 0 0 1 5.3-4.49c.4-.12.81.57 1.24 1.28.4.68.82 1.37 1.24 1.37.45 0 .9-.68 1.33-1.35.45-.7.89-1.38 1.32-1.25a8.61 8.61 0 0 1 5 4.17c3.73-2.94 5.1-7.74 5.1-10.7 0-2.34-1.57-1.6-4.09-.36l-.14.07c-2.31 1.15-5.39 2.67-8.77 2.67s-6.45-1.52-8.77-2.67c-2.6-1.29-4.23-2.1-4.23.29 0 3.05 1.46 8.06 5.47 10.97Z"+ clip-rule="evenodd"+ />+ <path+ fill="#FF9D0B"+ d="M70.71 37a3.25 3.25 0 1 0 0-6.5 3.25 3.25 0 0 0 0 6.5ZM24.21 37a3.25 3.25 0 1 0 0-6.5 3.25 3.25 0 0 0 0 6.5ZM17.52 48c-1.62 0-3.06.66-4.07 1.87a5.97 5.97 0 0 0-1.33 3.76 7.1 7.1 0 0 0-1.94-.3c-1.55 0-2.95.59-3.94 1.66a5.8 5.8 0 0 0-.8 7 5.3 5.3 0 0 0-1.79 2.82c-.24.9-.48 2.8.8 4.74a5.22 5.22 0 0 0-.37 5.02c1.02 2.32 3.57 4.14 8.52 6.1 3.07 1.22 5.89 2 5.91 2.01a44.33 44.33 0 0 0 10.93 1.6c5.86 0 10.05-1.8 12.46-5.34 3.88-5.69 3.33-10.9-1.7-15.92-2.77-2.78-4.62-6.87-5-7.77-.78-2.66-2.84-5.62-6.25-5.62a5.7 5.7 0 0 0-4.6 2.46c-1-1.26-1.98-2.25-2.86-2.82A7.4 7.4 0 0 0 17.52 48Zm0 4c.51 0 1.14.22 1.82.65 2.14 1.36 6.25 8.43 7.76 11.18.5.92 1.37 1.31 2.14 1.31 1.55 0 2.75-1.53.15-3.48-3.92-2.93-2.55-7.72-.68-8.01.08-.02.17-.02.24-.02 1.7 0 2.45 2.93 2.45 2.93s2.2 5.52 5.98 9.3c3.77 3.77 3.97 6.8 1.22 10.83-1.88 2.75-5.47 3.58-9.16 3.58-3.81 0-7.73-.9-9.92-1.46-.11-.03-13.45-3.8-11.76-7 .28-.54.75-.76 1.34-.76 2.38 0 6.7 3.54 8.57 3.54.41 0 .7-.17.83-.6.79-2.85-12.06-4.05-10.98-8.17.2-.73.71-1.02 1.44-1.02 3.14 0 10.2 5.53 11.68 5.53.11 0 .2-.03.24-.1.74-1.2.33-2.04-4.9-5.2-5.21-3.16-8.88-5.06-6.8-7.33.24-.26.58-.38 1-.38 3.17 0 10.66 6.82 10.66 6.82s2.02 2.1 3.25 2.1c.28 0 .52-.1.68-.38.86-1.46-8.06-8.22-8.56-11.01-.34-1.9.24-2.85 1.31-2.85Z"+ />+ <path+ fill="#FFD21E"+ d="M38.6 76.69c2.75-4.04 2.55-7.07-1.22-10.84-3.78-3.77-5.98-9.3-5.98-9.3s-.82-3.2-2.69-2.9c-1.87.3-3.24 5.08.68 8.01 3.91 2.93-.78 4.92-2.29 2.17-1.5-2.75-5.62-9.82-7.76-11.18-2.13-1.35-3.63-.6-3.13 2.2.5 2.79 9.43 9.55 8.56 11-.87 1.47-3.93-1.71-3.93-1.71s-9.57-8.71-11.66-6.44c-2.08 2.27 1.59 4.17 6.8 7.33 5.23 3.16 5.64 4 4.9 5.2-.75 1.2-12.28-8.53-13.36-4.4-1.08 4.11 11.77 5.3 10.98 8.15-.8 2.85-9.06-5.38-10.74-2.18-1.7 3.21 11.65 6.98 11.76 7.01 4.3 1.12 15.25 3.49 19.08-2.12Z"+ />+ <path+ fill="#FF9D0B"+ d="M77.4 48c1.62 0 3.07.66 4.07 1.87a5.97 5.97 0 0 1 1.33 3.76 7.1 7.1 0 0 1 1.95-.3c1.55 0 2.95.59 3.94 1.66a5.8 5.8 0 0 1 .8 7 5.3 5.3 0 0 1 1.78 2.82c.24.9.48 2.8-.8 4.74a5.22 5.22 0 0 1 .37 5.02c-1.02 2.32-3.57 4.14-8.51 6.1-3.08 1.22-5.9 2-5.92 2.01a44.33 44.33 0 0 1-10.93 1.6c-5.86 0-10.05-1.8-12.46-5.34-3.88-5.69-3.33-10.9 1.7-15.92 2.78-2.78 4.63-6.87 5.01-7.77.78-2.66 2.83-5.62 6.24-5.62a5.7 5.7 0 0 1 4.6 2.46c1-1.26 1.98-2.25 2.87-2.82A7.4 7.4 0 0 1 77.4 48Zm0 4c-.51 0-1.13.22-1.82.65-2.13 1.36-6.25 8.43-7.76 11.18a2.43 2.43 0 0 1-2.14 1.31c-1.54 0-2.75-1.53-.14-3.48 3.91-2.93 2.54-7.72.67-8.01a1.54 1.54 0 0 0-.24-.02c-1.7 0-2.45 2.93-2.45 2.93s-2.2 5.52-5.97 9.3c-3.78 3.77-3.98 6.8-1.22 10.83 1.87 2.75 5.47 3.58 9.15 3.58 3.82 0 7.73-.9 9.93-1.46.1-.03 13.45-3.8 11.76-7-.29-.54-.75-.76-1.34-.76-2.38 0-6.71 3.54-8.57 3.54-.42 0-.71-.17-.83-.6-.8-2.85 12.05-4.05 10.97-8.17-.19-.73-.7-1.02-1.44-1.02-3.14 0-10.2 5.53-11.68 5.53-.1 0-.19-.03-.23-.1-.74-1.2-.34-2.04 4.88-5.2 5.23-3.16 8.9-5.06 6.8-7.33-.23-.26-.57-.38-.98-.38-3.18 0-10.67 6.82-10.67 6.82s-2.02 2.1-3.24 2.1a.74.74 0 0 1-.68-.38c-.87-1.46 8.05-8.22 8.55-11.01.34-1.9-.24-2.85-1.31-2.85Z"+ />+ <path+ fill="#FFD21E"+ d="M56.33 76.69c-2.75-4.04-2.56-7.07 1.22-10.84 3.77-3.77 5.97-9.3 5.97-9.3s.82-3.2 2.7-2.9c1.86.3 3.23 5.08-.68 8.01-3.92 2.93.78 4.92 2.28 2.17 1.51-2.75 5.63-9.82 7.76-11.18 2.13-1.35 3.64-.6 3.13 2.2-.5 2.79-9.42 9.55-8.55 11 .86 1.47 3.92-1.71 3.92-1.71s9.58-8.71 11.66-6.44c2.08 2.27-1.58 4.17-6.8 7.33-5.23 3.16-5.63 4-4.9 5.2.75 1.2 12.28-8.53 13.36-4.4 1.08 4.11-11.76 5.3-10.97 8.15.8 2.85 9.05-5.38 10.74-2.18 1.69 3.21-11.65 6.98-11.76 7.01-4.31 1.12-15.26 3.49-19.08-2.12Z"+ />+</svg>diff --git a/docs/realistic-0.6b/index.html b/docs/realistic-0.6b/index.htmlnew file mode 100644index 0000000..dc98b79--- /dev/null+++ b/docs/realistic-0.6b/index.html@@ -0,0 +1,78 @@+<!doctype html>+<html lang="en">+<head>+ <meta charset="utf-8">+ <meta name="viewport" content="width=device-width, initial-scale=1">+ <meta name="color-scheme" content="light">+ <meta name="theme-color" content="#ffffff">+ <title>Audio8-TTS-Realistic · 0.6B — Listening Demo</title>+ <meta name="description" content="Listen to Audio8-TTS-Realistic: expressive speech in a compact 0.6B model, with 23 supported languages. Explore 24 examples alongside Preview, Breeze-TTS-2, Qwen3-TTS and Higgs TTS 3.">+ <meta property="og:type" content="website">+ <meta property="og:title" content="Audio8-TTS-Realistic · 0.6B">+ <meta property="og:description" content="More realistic speech. More languages. Still 0.6B. Explore 24 voice-cloning examples.">+ <meta property="og:url" content="https://edge0-ai.github.io/Audio8_TTS/realistic-0.6b/">+ <link rel="canonical" href="https://edge0-ai.github.io/Audio8_TTS/realistic-0.6b/">+ <link rel="stylesheet" href="../version-nav.css">+ <link rel="stylesheet" href="./styles.css?v=2">+</head>+<body>+ <a class="skip-link" href="#listening">Skip to listening examples</a>+ <header class="version-bar" id="top">+ <div class="shell version-inner">+ <a class="version-brand" href="../">Audio8 TTS</a>+ <nav class="version-switcher" aria-label="Model versions">+ <a href="../realistic-0.6b/" aria-current="page">Realistic <span>0.6B</span></a>+ <a href="../preview-0.6b/">Preview <span>0.6B</span></a>+ <a href="../0.1B/">Preview <span>0.1B</span></a>+ </nav>+ </div>+ </header>++ <main>+ <section class="hero shell" aria-labelledby="hero-title">+ <h1 id="hero-title">Audio8-TTS-Realistic</h1>+ <p class="hero-copy">More realistic speech.<br>More languages.<br>Still 0.6B.</p>+ <p class="hero-detail">Built on Audio8-TTS-Preview, with richer expression and support for <strong>23 languages</strong>.</p>+ <div class="hero-actions">+ <a class="button model-link" href="https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b"><img src="./icons/huggingface.svg" alt="" width="20" height="20">Hugging Face <span aria-hidden="true">↗</span></a>+ <a class="button github-link" href="https://github.com/Edge0-AI/Audio8_TTS"><img src="./icons/github.svg" alt="" width="18" height="18">GitHub <span aria-hidden="true">↗</span></a>+ <a class="button paper-link" href="https://arxiv.org/abs/2610.14776"><img src="./icons/arxiv.svg" alt="" width="18" height="18">Paper <span aria-hidden="true">↗</span></a>+ </div>+ <div class="hero-facts" aria-label="Model highlights">+ <span><strong>0.6B</strong> model parameters</span>+ <span><strong>23</strong> supported languages</span>+ <span><strong>12</strong> more than Preview</span>+ </div>+ </section>++ <section class="listening shell" id="listening" aria-labelledby="listening-title">+ <div class="section-heading">+ <div><p class="eyebrow">Listening room</p><h2 id="listening-title">A closer listen.</h2></div>+ <span class="example-total">24 examples · 5 models</span>+ </div>+ <p class="section-copy">The same text and reference voice, rendered by five models. Start with the reference and explore the samples at your own pace.</p>+ <p class="listening-note">Preview is included to illustrate the improvements in Realistic. The other models offer a broader reference for exploring different aspects of speech. Each has its own strengths; these selected examples are not an overall ranking of the models.</p>+ <div class="listening-controls">+ <div class="filters" role="group" aria-label="Filter examples by language">+ <button type="button" data-language="all" aria-pressed="true">All <span>24</span></button>+ <button type="button" data-language="en" aria-pressed="false">English <span>16</span></button>+ <button type="button" data-language="zh" aria-pressed="false">Chinese <span>8</span></button>+ </div>+ <label class="jump-label">Jump to <select id="sample-jump" aria-label="Jump to an example"><option value="">an example</option></select></label>+ </div>+ <p id="result-count" class="result-count" role="status">Showing all 24 examples</p>+ <div id="demo-root"><p class="loading">Loading listening examples…</p></div>+ <noscript><p>Enable JavaScript to load the listening examples and audio players.</p></noscript>+ </section>++ <aside class="notes shell" aria-labelledby="notes-title">+ <h2 id="notes-title">About these examples</h2>+ <p>These are curated English and Chinese voice-cloning examples. They illustrate speech realism and expression; they do not cover all 23 supported languages or represent a blind preference test. See the <a href="https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b#evaluation">model card</a> for language coverage and evaluation.</p>+ <p>Sizes use the projects’ published model-size conventions: Audio8 Realistic and Preview 0.6B, <a href="https://huggingface.co/BreezeBlue/Breeze-TTS-2">Breeze-TTS-2 ~3B</a>, <a href="https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base">Qwen3-TTS-12Hz-1.7B-Base 1.7B</a>, and <a href="https://huggingface.co/bosonai/higgs-tts-3-4b">Higgs TTS 3 ~4B</a>. Counting conventions differ; Audio8’s 0.6B excludes its separate audio codec.</p>+ </aside>+ </main>+ <footer class="site-footer shell"><span><strong>Audio8 TTS</strong> · Compact models, expressive speech.</span><a href="#top">Back to top ↑</a></footer>+ <script src="./data.js" defer></script>+ <script src="./app.js?v=2" defer></script>+</body>+</html>diff --git a/docs/realistic-0.6b/styles.css b/docs/realistic-0.6b/styles.cssnew file mode 100644index 0000000..df2ba63--- /dev/null+++ b/docs/realistic-0.6b/styles.css@@ -0,0 +1,156 @@+:root {+ --ink: #1d1d1f;+ --muted: #6e6e73;+ --line: #e5e5e9;+ --paper: #fff;+ --soft: #f5f5f7;+ --orange: #ff5a21;+ --navy: #2a2b3d;+ --sky: #3ab2da;+ --mint: #6fdba1;+}+* { box-sizing: border-box; }+html { scroll-behavior: smooth; scroll-padding-top: 24px; }+body { margin: 0; background: var(--paper); color: var(--ink); font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; font-size: 15px; line-height: 1.65; -webkit-font-smoothing: antialiased; }+a { color: #256e87; text-underline-offset: 3px; }+button, select { font: inherit; }+button, a, select { -webkit-tap-highlight-color: transparent; }+a:focus-visible, button:focus-visible, select:focus-visible, summary:focus-visible, audio:focus-visible { outline: 3px solid var(--sky); outline-offset: 4px; }+button { cursor: pointer; }+.shell { width: min(1200px, calc(100% - 64px)); margin-inline: auto; }+.skip-link { position: absolute; left: 18px; top: -80px; padding: 12px; background: #fff; z-index: 10; }+.skip-link:focus { top: 10px; }++/* Keep the quieter navigation local to the Realistic pages. */+.version-bar { background: #fff; border-color: var(--line); color: var(--ink); }+.version-inner { min-height: 76px; padding-block: 14px; }+.version-brand { color: var(--ink); font-size: 17px; font-weight: 600; letter-spacing: -.025em; }+.version-switcher { gap: 3px; padding: 4px; border: 0; border-radius: 12px; background: var(--soft); }+.version-switcher a { min-height: 34px; padding: 7px 12px; font-size: 12px; font-weight: 500; border-radius: 9px; color: var(--muted); }+.version-switcher a span { font-size: 10px; }+.version-switcher a:hover { color: var(--ink); background: #eaeaee; }+.version-switcher a[aria-current="page"] { color: var(--ink); background: #fff; box-shadow: 0 1px 4px #1d1d1f12; }+.version-bar a:focus-visible { outline-color: var(--sky); }++.hero { text-align: center; padding: 90px 0 72px; }+h1 { color: var(--muted); font-size: clamp(19px, 2.4vw, 24px); font-weight: 500; line-height: 1.3; letter-spacing: -.025em; margin: 0 0 22px; }+.hero-copy { color: var(--ink); font-size: clamp(35px, 5.8vw, 62px); font-weight: 600; line-height: 1.1; letter-spacing: -.045em; margin: 0 0 28px; }+.hero-detail { max-width: 520px; margin: 0 auto; color: var(--muted); font-size: 15px; }+.hero-detail strong { font-weight: 500; color: var(--ink); }+.hero-actions { display: flex; align-items: center; justify-content: center; flex-wrap: wrap; gap: 10px; margin: 28px 0 38px; }+.button { display: inline-flex; gap: 9px; align-items: center; justify-content: center; padding: 10px 17px; min-height: 43px; border: 1px solid var(--line); border-radius: 999px; background: #fff; color: var(--navy); font-size: 13px; font-weight: 500; text-decoration: none; transition: background-color .15s, border-color .15s; }+.button img { display: block; flex-shrink: 0; object-fit: contain; }+.button > span { color: var(--muted); font-size: 12px; margin-left: 4px; }+.button:hover { background: var(--soft); border-color: #c8c8ce; }+.model-link:hover { background: #fff5ef; border-color: var(--orange); }+.paper-link:hover { background: #f0faff; border-color: var(--sky); }+.hero-facts { display: flex; justify-content: center; gap: 0; color: var(--muted); font-size: 12px; }+.hero-facts span { padding: 0 30px; border-right: 1px solid var(--line); }+.hero-facts span:last-child { border: 0; }+.hero-facts strong { font-size: 20px; font-weight: 500; color: var(--ink); margin-right: 5px; font-variant-numeric: tabular-nums; }++.listening { padding-top: 50px; border-top: 1px solid var(--line); }+.section-heading { display: flex; align-items: end; justify-content: space-between; gap: 20px; }+.eyebrow { font-size: 12px; font-weight: 500; color: var(--muted); margin: 0 0 9px; }+h2 { font-size: 32px; line-height: 1.3; font-weight: 600; letter-spacing: -.035em; margin: 0; }+.example-total { font-size: 12px; color: var(--muted); white-space: nowrap; padding-bottom: 4px; }+.section-copy { font-size: 15px; color: var(--muted); max-width: 760px; margin: 15px 0 10px; line-height: 1.7; }+.listening-note { max-width: 850px; font-size: 12px; line-height: 1.8; color: var(--muted); margin: 0 0 30px; }+.listening-controls { display: flex; justify-content: space-between; align-items: center; gap: 16px; }+.filters { display: flex; flex-wrap: wrap; gap: 4px; background: var(--soft); padding: 4px; border-radius: 12px; }+.filters button { border: 1px solid transparent; border-radius: 9px; background: transparent; padding: 7px 14px; font-size: 12px; font-weight: 500; color: var(--muted); min-height: 36px; }+.filters button span { margin-left: 7px; font-weight: 400; opacity: .85; }+.filters button[aria-pressed=true] { background: #fff; color: var(--ink); box-shadow: 0 1px 4px #1d1d1f12; }+.filters button:hover { color: var(--ink); }+.jump-label { font-size: 12px; color: var(--muted); display: flex; gap: 8px; align-items: center; }+.jump-label select { border: 1px solid var(--line); background: #fff; border-radius: 9px; padding: 8px 28px 8px 10px; max-width: 190px; min-height: 38px; font-size: 12px; color: var(--ink); }+.result-count { font-size: 11px; color: var(--muted); margin: 13px 0 24px; }++.sample-card { background: #fff; border: 1px solid var(--line); border-radius: 18px; overflow: hidden; margin-bottom: 30px; scroll-margin-top: 24px; }+.sample-card[hidden] { display: none; }+.sample-head { padding: 26px 26px 21px; border-bottom: 1px solid #eeeef0; }+.sample-meta { display: flex; align-items: center; gap: 10px; margin-bottom: 15px; }+.sample-meta h3 { font-size: 12px; color: var(--muted); font-weight: 500; margin: 0; }+.language-tag { font-size: 10px; background: var(--soft); color: var(--muted); border-radius: 6px; padding: 2px 7px; }+.sample-link { margin-left: auto; color: #77777d; font-size: 13px; text-decoration: none; }+.prompt { font-size: 19px; line-height: 1.65; letter-spacing: -.02em; white-space: pre-line; margin: 0; max-width: 1080px; }+.prompt[lang=zh] { font-size: 19px; letter-spacing: .01em; }+.reference-text { margin-top: 15px; font-size: 12px; color: var(--muted); }+.reference-text summary { cursor: pointer; display: list-item; width: fit-content; }+.reference-text p { margin: 7px 0 0; max-width: 1000px; line-height: 1.7; }+.players { padding: 20px; display: grid; grid-template-columns: repeat(5, minmax(0, 1fr)); gap: 12px; }+.player { min-width: 0; background: #f7f7f9; border: 1px solid transparent; padding: 12px; border-radius: 12px; }+.player-heading { display: flex; flex-wrap: wrap; align-items: center; gap: 5px 7px; min-height: 43px; margin-bottom: 10px; align-content: center; }+.player h4 { margin: 0; font-size: 11px; line-height: 1.4; font-weight: 600; }+.size { display: inline-block; white-space: nowrap; font-size: 10px; font-weight: 500; padding: 1px 6px; background: #eaeaee; color: #53535a; border-radius: 5px; }+.player audio { width: 100%; display: block; height: 35px; min-width: 0; }+.duration { font-size: 10px; color: var(--muted); display: block; margin-top: 7px; font-variant-numeric: tabular-nums; }+.player.focus { grid-column: 1 / -1; background: #f7fbf8; border: 1px solid #d9e8df; border-left: 3px solid var(--mint); display: grid; grid-template-columns: minmax(200px, 1fr) minmax(250px, 2fr); align-items: center; gap: 20px; padding: 16px 18px; }+.focus .player-heading { margin: 0; min-height: 0; }+.focus h4 { color: var(--ink); font-size: 13px; }+.focus .size { background: #e8f3ec; color: #43594a; }+.focus-caption { display: block; width: 100%; font-size: 10px; color: var(--muted); margin-top: 2px; font-weight: 400; }+.audio-error { color: #a23434; font-size: 11px; margin: 4px 0 0; }+.notes { margin-top: 55px; padding: 32px 0 38px; border-top: 1px solid var(--line); }+.notes h2 { font-size: 15px; font-weight: 600; letter-spacing: 0; margin-bottom: 14px; }+.notes p { font-size: 12px; line-height: 1.8; color: var(--muted); max-width: 1040px; margin: 10px 0; }+.site-footer { border-top: 1px solid var(--line); padding: 25px 0 35px; display: flex; justify-content: space-between; gap: 18px; font-size: 11px; color: var(--muted); }+.site-footer strong { color: var(--ink); font-weight: 500; }+.site-footer a { text-decoration: none; }+.loading { padding: 50px; text-align: center; color: var(--muted); }++@media (max-width: 1100px) {+ .players { grid-template-columns: repeat(3, minmax(0, 1fr)); }+ .player-heading { min-height: 32px; }+ .player.focus { grid-template-columns: 1fr 2fr; }+ .shell { width: calc(100% - 40px); }+}+@media (max-width: 700px) {+ .shell { width: calc(100% - 32px); }+ .version-inner { gap: 12px; padding-block: 17px; }+ .version-switcher { width: 100%; }+ .version-switcher a { padding-inline: 7px; font-size: 11px; }+ .hero { padding: 62px 0 48px; }+ h1 { font-size: 19px; margin-bottom: 21px; }+ .hero-copy { font-size: clamp(31px, 7.1vw, 45px); margin-bottom: 24px; }+ .hero-detail { font-size: 13px; max-width: 360px; }+ .hero-actions { margin: 24px 0 29px; gap: 7px; }+ .button { font-size: 12px; padding: 9px 12px; gap: 7px; }+ .button > span { margin-left: 0; }+ .hero-facts { font-size: 10px; }+ .hero-facts span { flex: 1; padding: 0 9px; max-width: 150px; }+ .hero-facts strong { display: block; font-size: 22px; margin: 0 0 3px; }+ .listening { padding-top: 36px; }+ .section-heading { align-items: end; gap: 12px; }+ .section-heading h2 { font-size: 28px; }+ .example-total { font-size: 10px; white-space: normal; text-align: right; max-width: 95px; padding-bottom: 3px; }+ .section-copy { font-size: 14px; }+ .listening-note { margin-bottom: 25px; font-size: 11px; }+ .listening-controls { align-items: start; flex-direction: column; gap: 14px; }+ .jump-label select { max-width: 240px; }+ .sample-head { padding: 21px 17px 18px; }+ .prompt { font-size: 16px; }+ .prompt[lang=zh] { font-size: 17px; }+ .players { padding: 12px; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 9px; }+ .player { padding: 11px; }+ .player.reference { grid-column: 1 / -1; display: grid; grid-template-columns: 1fr 2fr; gap: 10px; align-items: center; }+ .reference .player-heading { margin: 0; }+ .player.focus { display: block; padding: 14px; }+ .focus .player-heading { margin-bottom: 12px; }+ .focus-caption { width: auto; margin: 0; }+ .player-heading { min-height: 39px; }+ .site-footer { align-items: start; font-size: 10px; }+ .site-footer span { max-width: 230px; }+}+@media (max-width: 380px) {+ .players { grid-template-columns: 1fr; }+ .player.reference { display: block; }+ .reference .player-heading { margin-bottom: 9px; }+ .player-heading { min-height: 0; }+ .button { padding: 9px 10px; }+ .button > span { display: none; }+}+@media (prefers-reduced-motion: reduce) {+ html { scroll-behavior: auto; }+ .button { transition: none; }+}diff --git a/docs/version-nav.css b/docs/version-nav.cssnew file mode 100644index 0000000..9c8d123--- /dev/null+++ b/docs/version-nav.css@@ -0,0 +1,73 @@+/* Shared navigation for the preserved Audio8 listening pages. */+.version-bar {+ color: #253b38;+ background: #fff;+ border-bottom: 1px solid #dfe9e6;+ font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif;+}+.version-inner {+ display: flex;+ min-height: 78px;+ align-items: center;+ justify-content: space-between;+ gap: 20px;+ padding-block: 15px;+}+.version-brand {+ color: #183b35;+ font-size: 18px;+ font-weight: 750;+ text-decoration: none;+ white-space: nowrap;+}+.version-brand span {+ display: block;+ color: #71817d;+ font-size: 11px;+ font-weight: 500;+ letter-spacing: .02em;+}+.version-switcher {+ display: flex;+ flex-wrap: wrap;+ gap: 5px;+ padding: 5px;+ border: 1px solid #e2ebe8;+ border-radius: 14px;+ background: #f5f8f7;+}+.version-switcher a {+ display: inline-flex;+ align-items: center;+ gap: 7px;+ min-height: 38px;+ padding: 8px 13px;+ border-radius: 10px;+ color: #52645e;+ font-size: 13px;+ font-weight: 650;+ text-decoration: none;+ white-space: nowrap;+}+.version-switcher a span { font-size: 11px; opacity: .8; }+.version-switcher a:hover { color: #087f72; background: #e7f0ed; }+.version-switcher a[aria-current="page"] { color: #fff; background: #087f72; }+.version-bar a:focus-visible, .version-resources a:focus-visible {+ outline: 3px solid #087f72;+ outline-offset: 4px;+}+.version-resources { display: flex; flex-wrap: wrap; gap: 8px; }+.version-resources .repo-link { border-radius: 8px; }+.version-resources .repo-link:first-child { background: #087f72; }+.version-resources .repo-link:first-child:hover { background: #06665b; }+@media (max-width: 680px) {+ .version-inner { align-items: flex-start; flex-direction: column; gap: 12px; }+ .version-brand span { display: inline; margin-left: 9px; }+ .version-switcher { width: 100%; }+ .version-switcher a { flex: 1; justify-content: center; padding-inline: 8px; }+ .version-resources { margin-top: 16px; }+ .version-resources .repo-link { margin-top: 0; }+}+@media (max-width: 360px) {+ .version-switcher a { flex-direction: column; gap: 0; font-size: 12px; }+}diff --git a/guides/deployment.md b/guides/deployment.mdnew file mode 100644index 0000000..fe80deb--- /dev/null+++ b/guides/deployment.md@@ -0,0 +1,252 @@+# Deployment guide++[All releases](../README.md) · [中文](deployment_zh.md) · [Installation and inference](inference.md)++| Release | Transformers / CLI | SGLang Omni | ONNX |+|---|---|---|---|+| Realistic 0.6B | Supported | 0.6B configuration | No export available |+| Preview 0.6B | Supported | 0.6B configuration | INT4, onnx_runtime/ |+| Preview 0.1B | Supported | 0.1B hybrid configuration | INT8, onnx_runtime_0_1b_int8/ |++Realistic uses the 0.6B configuration with its own explicit model ID or snapshot path. ONNX exports belong to the listed release; the Preview ONNX files are not Realistic weights.++## SGLang Omni Serving++The adapter in [`sglang_omni/`](../sglang_omni/) provides an OpenAI-compatible+service with SGLang paged attention, dynamic batching, a fixed KV cache for the+fast codebook decoder, reference-audio encoding, and waveform decoding. It is+installed as an independent `audio8_tts` model plugin and does not overwrite+SGLang Omni core files.++### Compatibility++The adapter uses internal SGLang Omni interfaces, so deploy it with the tested+revision instead of the latest `main` branch.++| Dependency | Tested version |+|---|---|+| SGLang Omni | `68a572348837f7b004857b4b07993c20ade4c017` (`0.1.0`) |+| SGLang | `0.5.8` |+| PyTorch | `2.9.1+cu128` |+| Transformers | `4.57.1` |+| Precision | BF16 |++### Performance++The existing SGLang benchmark and its original test conditions are retained below. Realistic and Preview 0.6B share the same architecture and serving path. These are existing measurements, not a new Realistic benchmark. The separate 0.1B eager hybrid path is described below.++Warm single-stream latency was measured on one NVIDIA H20 with BF16 weights,+CUDA Graph, greedy decoding, and 128 generated frames. The output WAV was+5.85-5.94 seconds long; cold start and compilation time were excluded. Lower+RTF is better.++| SGLang Omni adapter | Warm p50 latency | RTF |+|---|---:|---:|+| Current implementation | **0.691 s** | **0.116** |++See the [SGLang Omni implementation and evaluation report](../sglang_omni/OPTIMIZATION_REPORT.md)+for the configuration, implementation details, and validation results.++### Install++Run these commands from the Audio8 TTS repository root. The example uses+Python 3.12 and [`uv`](https://docs.astral.sh/uv/).++```bash+export SGLANG_OMNI_ROOT=/opt/sglang-omni+export MODEL_ID=Edge0/Audio8-TTS-Realistic-0.6b+# Preview 0.6B: export MODEL_ID=Edge0/Audio8-TTS-Preview-0.6b+export MODEL="./model/${MODEL_ID##*/}"++git clone https://github.com/sgl-project/sglang-omni.git "${SGLANG_OMNI_ROOT}"+git -C "${SGLANG_OMNI_ROOT}" checkout 68a572348837f7b004857b4b07993c20ade4c017++uv venv .venv-sglang --python 3.12+source .venv-sglang/bin/activate+uv pip install -v -e "${SGLANG_OMNI_ROOT}"++hf download "${MODEL_ID}" --local-dir "${MODEL}"+./sglang_omni/scripts/install_adapter.sh "${SGLANG_OMNI_ROOT}"+python3 ./sglang_omni/scripts/verify_install.py --model-path "${MODEL}"+```++For an existing wheel or site-packages installation, resolve the package+directory and install the adapter there:++```bash+SGLANG_OMNI_PACKAGE="$(python3 -c 'import importlib.util, pathlib; s=importlib.util.find_spec("sglang_omni"); assert s and s.origin; print(pathlib.Path(s.origin).parent)')"+./sglang_omni/scripts/install_adapter.sh "${SGLANG_OMNI_PACKAGE}"+```++### Start the service++```bash+CUDA_VISIBLE_DEVICES=0 \+SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \+MODEL="${MODEL}" \+AUDIO8_TTS_ENABLE_TORCH_COMPILE=1 \+HOST=0.0.0.0 \+PORT=8010 \+./sglang_omni/scripts/run_server.sh+```++The default `fa3` attention backend is intended for Hopper GPUs such as H20+and H100. Consumer Blackwell GPUs such as RTX 5090 report compute capability+`(12, 0)` and have no FA3 kernel image, so the adapter detects them and selects+FlashInfer for the SGLang slow-AR path automatically; the short fixed-cache+fast head then uses PyTorch SDPA. No configuration is required.++Setting the variable explicitly still overrides the detection on any GPU:++```bash+AUDIO8_TTS_ATTENTION_BACKEND=flashinfer \+CUDA_VISIBLE_DEVICES=0 \+SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \+MODEL="${MODEL}" \+./sglang_omni/scripts/run_server.sh+```++The defaults use model name `audio8/tts-0.6b`, BF16, one GPU, a `0.2` static+memory fraction, and up to 32 running requests. The main runtime controls are+`MODEL_NAME`, `AUDIO8_TTS_MEM_FRACTION_STATIC`,+`AUDIO8_TTS_MAX_RUNNING_REQUESTS`, `AUDIO8_TTS_CHUNKED_PREFILL_SIZE`, and+`AUDIO8_TTS_DISABLE_CUDA_GRAPH`. When Torch compilation is enabled, the adapter+uses SGLang's native batch-size policy. Set `AUDIO8_TTS_TORCH_COMPILE_MAX_BS`+only when an explicit compile limit is needed. `AUDIO8_TTS_ATTENTION_BACKEND`+defaults to `fa3`, except on GPUs with no FA3 kernel image such as consumer+Blackwell, where it defaults to `flashinfer`; set it explicitly to override. Set+`SGLANG_OMNI_SITE_PACKAGES` when the runtime dependencies are installed in a+separate site-packages directory.++### Troubleshooting++- Install the distribution package that provides `libnuma` if `sgl_kernel`+ fails to import (for example, `numactl` or `libnuma1`).+- Put the CUDA toolkit `bin` directory on `PATH`. If `deep_gemm` cannot find+ `nvcc` during JIT compilation, also set `CUDA_PATH` to the toolkit root.+- Keep Transformers on the supported 4.x range (`>=4.57.0,<5`). Transformers+ 5.x can produce invalid all-zero codes for this custom-code model.++### Call the API++Generate speech without a reference:++```bash+curl -sS --fail-with-body \+ -H 'Content-Type: application/json' \+ -d '{+ "model": "audio8/tts-0.6b",+ "input": "Hello from Audio8 TTS.",+ "response_format": "wav",+ "max_new_tokens": 256,+ "temperature": 0.8,+ "top_p": 0.95,+ "top_k": 50+ }' \+ http://127.0.0.1:8010/v1/audio/speech \+ -o audio8.wav+```++Generate speech with one reference voice:++```bash+curl -sS --fail-with-body \+ -H 'Content-Type: application/json' \+ -d '{+ "model": "audio8/tts-0.6b",+ "input": "This sentence uses the reference voice.",+ "response_format": "wav",+ "temperature": 0.8,+ "top_p": 0.95,+ "top_k": 50,+ "references": [{+ "audio_path": "/data/reference.wav",+ "text": "The exact transcript of the reference recording."+ }]+ }' \+ http://127.0.0.1:8010/v1/audio/speech \+ -o audio8_clone.wav+```++The reference path must be visible inside the service environment. The current+adapter supports TP=1 and one reference per request. Set `"stream": true` to+receive SSE audio chunks as they are generated. For the lowest overhead, use+`"response_format": "pcm"`; each event contains Base64-encoded audio in+`audio.data`. Streaming defaults to 12 codec frames per chunk, with 128 frames+of decoder context and a one-frame boundary guard. Override these values with+`AUDIO8_TTS_STREAM_CHUNK_FRAMES`, `AUDIO8_TTS_STREAM_CONTEXT_FRAMES`, and+`AUDIO8_TTS_STREAM_GUARD_FRAMES`.++Streaming is a server-side opt-in. Start the service with+`AUDIO8_TTS_STREAM_ENABLED=1` to enable SSE streaming; the default is off, and+requests are answered with the complete audio after generation finishes+(a request-level `"stream": true` is then ignored). Non-streaming remains the+default for maximum throughput and stable memory usage. `response_format` of+`codes`/`codec`/`npy` always returns codec codes without streaming.++Run the smoke test to verify a deployment:++```bash+BASE_URL=http://127.0.0.1:8010 ./sglang_omni/scripts/smoke_test.sh+python3 ./sglang_omni/scripts/stream_smoke_test.py \+ --base-url http://127.0.0.1:8010 \+ --output /tmp/audio8_stream.wav+```++To build the adapter into an existing image, append+[`sglang_omni/Dockerfile.snippet`](../sglang_omni/Dockerfile.snippet) after the+SGLang Omni package and its Python dependencies are installed.++### Audio8-TTS-0.1B (Falcon-H1 hybrid slow backbone)++Install the adapter as above, then select this model, configuration, and API name. Use audio8/tts-0.1b in the request's model field.++`Audio8-TTS-Preview-0.1b` replaces the pure-attention slow AR backbone with a+Falcon-H1 hybrid (Mamba 2 SSM + attention). The adapter detects this+automatically from `config.json` (`slow_backbone: falcon_h1` or a `mamba_d_ssm`+field) and switches to the eager hybrid path: the slow backbone runs with one+`FalconHybridMambaAttentionDynamicCache` per request, while the fast codebook+AR, semantic sampling and vocoder are shared with the 0.6B path. The 0.6B+path is unchanged.++```bash+export MODEL=./model/Audio8-TTS-Preview-0.1b+hf download Edge0/Audio8-TTS-Preview-0.1b --local-dir "${MODEL}"+export CONFIG=./sglang_omni/configs/audio8_tts_0_1b.yaml+export MODEL_NAME=audio8/tts-0.1b++CUDA_VISIBLE_DEVICES=0 \+SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \+MODEL="${MODEL}" \+CONFIG="${CONFIG}" \+MODEL_NAME="${MODEL_NAME}" \+./sglang_omni/scripts/run_server.sh+```++Because the Mamba backbone keeps its own hybrid caches and does not consume+SGLang KV pages, the adapter disables CUDA graphs and caps the static memory+fraction for this model. The slow backbone is bit-exact with the Transformers+Falcon-H1 implementation for both prefill and cached decode.++## Preview 0.6B: CPU ONNX INT4++[`onnx_runtime/`](../onnx_runtime/) provides a standalone CPU deployment using+weight-only INT4 Slow/Fast AR models, FP16 activations and KV caches, and an+FP16 codec. It includes CLI inference, a local web and HTTP service, streaming+PCM output, and reference-voice registration without PyTorch or Transformers.++The online sessions use about 1 GiB of memory in the tested Apple M2 setup.+During voice registration, the online sessions are released before the codec+encoder is loaded to keep peak memory low.++Download the ONNX model from+[Audio8-TTS-Preview-0.6B-ONNX-INT4](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6B-ONNX-INT4)+and follow the [ONNX Runtime guide](../onnx_runtime/README.md).++## Preview 0.1B: CPU ONNX INT8++0.1B uses separate Falcon-H1 INT8 graphs and a dedicated runtime; they are incompatible with the 0.6B INT4 graphs. It includes CLI, web/HTTP serving, streaming PCM and voice registration, with Linux/macOS and Windows PowerShell entry points.++[Model](https://huggingface.co/Edge0/audio8-TTS-0.1B-ONNX-INT8) · [Runtime guide](../onnx_runtime_0_1b_int8/README.md)++Consult each download repository for its model/export license.diff --git a/guides/deployment_zh.md b/guides/deployment_zh.mdnew file mode 100644index 0000000..e55d94d--- /dev/null+++ b/guides/deployment_zh.md@@ -0,0 +1,235 @@+# 部署指南++[系列首页](../README_zh.md) · [English](deployment.md) · [安装与推理](inference_zh.md)++| 版本 | Transformers / CLI | SGLang Omni | ONNX |+|---|---|---|---|+| Realistic 0.6B | 支持 | 0.6B 配置 | 暂无制品 |+| Preview 0.6B | 支持 | 0.6B 配置 | INT4,onnx_runtime/ |+| Preview 0.1B | 支持 | 0.1B hybrid 配置 | INT8,onnx_runtime_0_1b_int8/ |++Realistic 沿用 0.6B 配置,只需显式选择新版模型 ID 或本地目录。ONNX 导出属于表中列出的各自版本,不能将 Preview ONNX 当成 Realistic 权重。++## SGLang Omni 服务部署++[`sglang_omni/`](../sglang_omni/) 中的适配器提供兼容 OpenAI API 的推理服务,包含 SGLang+paged attention、动态 batching、Fast AR 固定 KV cache、参考音频编码和 waveform 解码。+它以独立的 `audio8_tts` 模型插件安装,不会覆盖 SGLang Omni 核心文件。++### 兼容版本++适配器使用了 SGLang Omni 的内部接口,因此部署时应固定到以下已验证版本,不要直接使用最新+`main` 分支。++| 依赖 | 已验证版本 |+|---|---|+| SGLang Omni | `68a572348837f7b004857b4b07993c20ade4c017`(`0.1.0`) |+| SGLang | `0.5.8` |+| PyTorch | `2.9.1+cu128` |+| Transformers | `4.57.1` |+| 精度 | BF16 |++### 性能++现有 SGLang 性能数据与原测试条件如下;Realistic 与 Preview 0.6B 架构和服务路径相同,继续适用。此处不将其写作新版重新测量的数据。0.1B 的 eager 混合路径见本页独立小节。++单流 warm latency 在单张 NVIDIA H20 上测试,使用 BF16、CUDA Graph、greedy decoding,+并生成 128 帧。输出 WAV 时长为 5.85-5.94 秒,不包含冷启动和编译时间。RTF 越低越好。++| SGLang Omni 适配器 | Warm p50 latency | RTF |+|---|---:|---:|+| 当前实现 | **0.691 s** | **0.116** |++配置、实现细节和验证结果请参阅+[SGLang Omni 实现与评测报告](../sglang_omni/OPTIMIZATION_REPORT.md)。++### 安装++在 Audio8 TTS 仓库根目录执行以下命令。示例使用 Python 3.12 和+[`uv`](https://docs.astral.sh/uv/)。++```bash+export SGLANG_OMNI_ROOT=/opt/sglang-omni+export MODEL_ID=Edge0/Audio8-TTS-Realistic-0.6b+# Preview 0.6B: export MODEL_ID=Edge0/Audio8-TTS-Preview-0.6b+export MODEL="./model/${MODEL_ID##*/}"++git clone https://github.com/sgl-project/sglang-omni.git "${SGLANG_OMNI_ROOT}"+git -C "${SGLANG_OMNI_ROOT}" checkout 68a572348837f7b004857b4b07993c20ade4c017++uv venv .venv-sglang --python 3.12+source .venv-sglang/bin/activate+uv pip install -v -e "${SGLANG_OMNI_ROOT}"++hf download "${MODEL_ID}" --local-dir "${MODEL}"+./sglang_omni/scripts/install_adapter.sh "${SGLANG_OMNI_ROOT}"+python3 ./sglang_omni/scripts/verify_install.py --model-path "${MODEL}"+```++如果 SGLang Omni 已通过 wheel 或 site-packages 安装,可以定位其包目录后安装适配器:++```bash+SGLANG_OMNI_PACKAGE="$(python3 -c 'import importlib.util, pathlib; s=importlib.util.find_spec("sglang_omni"); assert s and s.origin; print(pathlib.Path(s.origin).parent)')"+./sglang_omni/scripts/install_adapter.sh "${SGLANG_OMNI_PACKAGE}"+```++### 启动服务++```bash+CUDA_VISIBLE_DEVICES=0 \+SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \+MODEL="${MODEL}" \+AUDIO8_TTS_ENABLE_TORCH_COMPILE=1 \+HOST=0.0.0.0 \+PORT=8010 \+./sglang_omni/scripts/run_server.sh+```++默认的 `fa3` attention backend 适用于 H20、H100 等 Hopper GPU。RTX 5090 等消费级+Blackwell GPU 的计算能力为 `(12, 0)`,没有对应的 FA3 kernel image,适配器会自动识别+并为 SGLang Slow AR 路径选择 FlashInfer;此时短序列固定 KV cache 的 Fast head 会使用+PyTorch SDPA。无需额外配置。++在任意 GPU 上显式设置该环境变量仍会覆盖自动检测结果:++```bash+AUDIO8_TTS_ATTENTION_BACKEND=flashinfer \+CUDA_VISIBLE_DEVICES=0 \+SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \+MODEL="${MODEL}" \+./sglang_omni/scripts/run_server.sh+```++默认配置使用模型名 `audio8/tts-0.6b`、BF16、单卡、`0.2` 静态显存比例和最多 32 个并发+请求。主要运行参数包括 `MODEL_NAME`、`AUDIO8_TTS_MEM_FRACTION_STATIC`、+`AUDIO8_TTS_MAX_RUNNING_REQUESTS`、`AUDIO8_TTS_CHUNKED_PREFILL_SIZE` 和+`AUDIO8_TTS_DISABLE_CUDA_GRAPH`。开启 Torch compile 后,适配器沿用 SGLang 原生的 batch+策略;只有需要明确限制 compile 上限时才设置 `AUDIO8_TTS_TORCH_COMPILE_MAX_BS`。+`AUDIO8_TTS_ATTENTION_BACKEND` 默认为 `fa3`;在没有 FA3 kernel image 的 GPU(如消费级+Blackwell)上默认为 `flashinfer`,显式设置可覆盖。如果运行依赖安装在单独的+site-packages 目录中,请设置+`SGLANG_OMNI_SITE_PACKAGES`。++### 常见问题++- 如果 `sgl_kernel` 导入失败,请安装系统中提供 `libnuma` 的软件包,例如 `numactl` 或+ `libnuma1`。+- CUDA toolkit 的 `bin` 目录需要加入 `PATH`。如果 `deep_gemm` JIT 编译时找不到+ `nvcc`,还需要将 `CUDA_PATH` 指向 CUDA toolkit 根目录。+- Transformers 应保持在支持的 4.x 范围(`>=4.57.0,<5`)。Transformers 5.x 可能使该+ custom-code 模型生成无效的全零 codes。++### 调用 API++无参考音频生成:++```bash+curl -sS --fail-with-body \+ -H 'Content-Type: application/json' \+ -d '{+ "model": "audio8/tts-0.6b",+ "input": "你好,这是 Audio8 TTS 服务测试。",+ "response_format": "wav",+ "max_new_tokens": 256,+ "temperature": 0.8,+ "top_p": 0.95,+ "top_k": 50+ }' \+ http://127.0.0.1:8010/v1/audio/speech \+ -o audio8.wav+```++使用一个参考音色生成:++```bash+curl -sS --fail-with-body \+ -H 'Content-Type: application/json' \+ -d '{+ "model": "audio8/tts-0.6b",+ "input": "这句话使用参考音色生成。",+ "response_format": "wav",+ "temperature": 0.8,+ "top_p": 0.95,+ "top_k": 50,+ "references": [{+ "audio_path": "/data/reference.wav",+ "text": "参考录音对应的准确文本。"+ }]+ }' \+ http://127.0.0.1:8010/v1/audio/speech \+ -o audio8_clone.wav+```++参考音频路径必须在服务运行环境中可见。当前适配器支持 TP=1 和每个请求一个 reference。+设置 `"stream": true` 可在生成过程中接收 SSE 音频分块。使用+`"response_format": "pcm"` 开销最低,每个事件的 `audio.data` 包含 Base64 编码音频。+流式输出默认每 12 个 codec frames 生成一块,使用 128 帧 decoder context 和 1 帧边界+guard。可通过 `AUDIO8_TTS_STREAM_CHUNK_FRAMES`、+`AUDIO8_TTS_STREAM_CONTEXT_FRAMES` 和 `AUDIO8_TTS_STREAM_GUARD_FRAMES` 调整。++流式输出是服务端可选项:启动服务时设置 `AUDIO8_TTS_STREAM_ENABLED=1` 才会启用 SSE+流式;默认关闭,请求会在生成完成后一次性返回完整音频(请求中的 `"stream": true`+此时会被忽略)。默认非流式以获得更高的吞吐和稳定的显存占用。+`response_format` 为 `codes`/`codec`/`npy` 时始终非流式,直接返回 codec codes。++运行以下 smoke test 可以验证部署:++```bash+BASE_URL=http://127.0.0.1:8010 ./sglang_omni/scripts/smoke_test.sh+python3 ./sglang_omni/scripts/stream_smoke_test.py \+ --base-url http://127.0.0.1:8010 \+ --output /tmp/audio8_stream.wav+```++构建镜像时,请在安装 SGLang Omni 包及其 Python 依赖后追加+[`sglang_omni/Dockerfile.snippet`](../sglang_omni/Dockerfile.snippet)。++### Audio8-TTS-0.1B(Falcon-H1 混合慢速主干)++先完成前述适配器安装,再选择下面的模型、配置和 API 名称。请求中的 model 字段应使用 audio8/tts-0.1b。++`Audio8-TTS-Preview-0.1b` 将纯注意力的慢速 AR 主干替换为 Falcon-H1 混合结构+(Mamba 2 SSM + 注意力)。适配器会自动从 `config.json` 检测(`slow_backbone:+falcon_h1` 或存在 `mamba_d_ssm` 字段)并切换到 eager 混合路径:慢速主干为每个+请求维护一个 `FalconHybridMambaAttentionDynamicCache`,而快速码本 AR、语义采样+与声码器与 0.6B 路径完全共用,0.6B 路径不受影响。++```bash+export MODEL=./model/Audio8-TTS-Preview-0.1b+hf download Edge0/Audio8-TTS-Preview-0.1b --local-dir "${MODEL}"+export CONFIG=./sglang_omni/configs/audio8_tts_0_1b.yaml+export MODEL_NAME=audio8/tts-0.1b++CUDA_VISIBLE_DEVICES=0 \+SGLANG_OMNI_ROOT="${SGLANG_OMNI_ROOT}" \+MODEL="${MODEL}" \+CONFIG="${CONFIG}" \+MODEL_NAME="${MODEL_NAME}" \+./sglang_omni/scripts/run_server.sh+```++由于 Mamba 主干自行维护混合缓存、不占用 SGLang KV 页,适配器会为该模型关闭+CUDA Graph 并限制静态显存比例。慢速主干在 prefill 与增量 decode 上与+Transformers 的 Falcon-H1 实现逐位一致。++## Preview 0.6B: CPU ONNX INT4++[`onnx_runtime/`](../onnx_runtime/) 提供独立的 CPU 部署方案:Slow/Fast AR 使用+weight-only INT4,activation 和 KV cache 使用 FP16,codec 使用 FP16。它包含+命令行推理、本地网页与 HTTP 服务、流式 PCM 输出和参考音频音色注册,运行时不依赖+PyTorch 或 Transformers。++在 Apple M2 测试环境中,在线模型 session 约占 1 GiB 内存。注册音色前会先释放+在线 session,再单独加载 codec encoder,从而控制峰值内存。++从+[Audio8-TTS-Preview-0.6B-ONNX-INT4](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6B-ONNX-INT4)+下载模型,并参阅 [ONNX Runtime 中文指南](../onnx_runtime/README_zh.md)。++## Preview 0.1B: CPU ONNX INT8++0.1B 使用独立的 Falcon-H1 INT8 图及运行时;与 0.6B INT4 图不兼容。提供 CLI、网页/HTTP 服务、流式 PCM 和音色注册,支持 Linux/macOS 以及 Windows PowerShell 入口。++[Model](https://huggingface.co/Edge0/audio8-TTS-0.1B-ONNX-INT8) · [运行时指南](../onnx_runtime_0_1b_int8/README_zh.md)++模型权重及 ONNX 制品的许可请以各下载仓库为准。diff --git a/guides/inference.md b/guides/inference.mdnew file mode 100644index 0000000..70aa3d6--- /dev/null+++ b/guides/inference.md@@ -0,0 +1,80 @@+# Installation and inference++[All releases](../README.md) · [中文](inference_zh.md) · [Deployment](deployment.md) · [SFT](training.md)++## Installation++Run from the repository root. Python 3.10+ and a CUDA GPU are recommended for 0.6B; the 0.1B model card recommends Python 3.11+.++```bash+python3 -m venv .venv+source .venv/bin/activate+pip install -r requirements.txt+```++## Choose a release++Choose one model ID below and set MODEL. Every following command uses that release. A complete local checkpoint directory is also accepted.++| Release | MODEL |+|---|---|+| Realistic 0.6B | `Edge0/Audio8-TTS-Realistic-0.6b` |+| Preview 0.6B | `Edge0/Audio8-TTS-Preview-0.6b` |+| Preview 0.1B | `Edge0/Audio8-TTS-Preview-0.1b` |++```bash+export MODEL=Edge0/Audio8-TTS-Realistic-0.6b+# Preview 0.6B: export MODEL=Edge0/Audio8-TTS-Preview-0.6b+# Preview 0.1B: export MODEL=Edge0/Audio8-TTS-Preview-0.1b+```++If you arrived from a Preview guide, select its Preview ID. Without --model, the scripts still default to model/audio8_tts_0_6B_preview/; this guide always passes the model explicitly.++## Inference++For best synthesis quality, keep each input within 150 characters. Longer text+may reduce generation quality; split it into shorter segments when needed.++### Zero-shot voice cloning++Use your own recording at `reference.wav`. Its transcript should match the spoken content exactly.++```bash+python audio8_tts_infer.py \+ --model "${MODEL}" \+ --text "Welcome to audio8_tts." \+ --reference-audio reference.wav \+ --reference-text "Transcript of the reference recording." \+ --output outputs/clone.wav+```++### Generation without a reference++```bash+python audio8_tts_infer.py \+ --model "${MODEL}" \+ --text "This utterance does not use a reference voice." \+ --output outputs/no_reference.wav+```++### Batch inference++Each line in the input manifest is an independent JSON object. Relative audio+paths are resolved from the manifest directory.++```json+{"id":"sample_001","text":"Target text","reference_audio":"audio/ref.wav","reference_text":"Reference transcript"}+{"id":"sample_002","text":"Text without a reference voice"}+```++```bash+python audio8_tts_infer.py \+ --model "${MODEL}" \+ --input-jsonl data/prompts.jsonl \+ --output-dir outputs/batch \+ --batch-size 2+```++The batch command writes `manifest.jsonl` and `failures.jsonl`. Existing WAV+files are skipped unless `--overwrite` is passed. See+`python audio8_tts_infer.py --help` for sampling and code-saving options.diff --git a/guides/inference_zh.md b/guides/inference_zh.mdnew file mode 100644index 0000000..21e0401--- /dev/null+++ b/guides/inference_zh.md@@ -0,0 +1,78 @@+# 安装与推理++[系列首页](../README_zh.md) · [English](inference.md) · [部署](deployment_zh.md) · [SFT](training_zh.md)++## 安装++在仓库根目录执行。0.6B 推荐 Python 3.10+ 与 CUDA GPU;0.1B 按其模型卡推荐 Python 3.11+。++```bash+python3 -m venv .venv+source .venv/bin/activate+pip install -r requirements.txt+```++## 选择版本++从下表选择一个模型 ID,设为 MODEL;后续全部命令会使用该版本。也可以填入完整的本地模型目录。++| 版本 | MODEL |+|---|---|+| Realistic 0.6B | `Edge0/Audio8-TTS-Realistic-0.6b` |+| Preview 0.6B | `Edge0/Audio8-TTS-Preview-0.6b` |+| Preview 0.1B | `Edge0/Audio8-TTS-Preview-0.1b` |++```bash+export MODEL=Edge0/Audio8-TTS-Realistic-0.6b+# Preview 0.6B: export MODEL=Edge0/Audio8-TTS-Preview-0.6b+# Preview 0.1B: export MODEL=Edge0/Audio8-TTS-Preview-0.1b+```++若从 Preview 指南进入,请选择相应 Preview ID。脚本不传 --model 时仍默认读取 model/audio8_tts_0_6B_preview/,本指南始终显式传参。++## 推理++为获得更好的合成效果,建议单次输入不超过 150 个字。文本过长可能影响生成质量,请根据+需要拆分为较短的片段。++### 零样本音色克隆++将自己的参考录音保存为 `reference.wav`,参考文本应与录音实际内容一致。++```bash+python audio8_tts_infer.py \+ --model "${MODEL}" \+ --text "欢迎使用 audio8_tts。" \+ --reference-audio reference.wav \+ --reference-text "参考录音对应的准确文本。" \+ --output outputs/clone.wav+```++### 无参考音频生成++```bash+python audio8_tts_infer.py \+ --model "${MODEL}" \+ --text "这条语音不使用参考音色。" \+ --output outputs/no_reference.wav+```++### 批量推理++输入文件每行是一个独立 JSON 对象。相对音频路径相对于 JSONL 文件所在目录解析。++```json+{"id":"sample_001","text":"目标文本","reference_audio":"audio/ref.wav","reference_text":"参考文本"}+{"id":"sample_002","text":"不使用参考音色的文本"}+```++```bash+python audio8_tts_infer.py \+ --model "${MODEL}" \+ --input-jsonl data/prompts.jsonl \+ --output-dir outputs/batch \+ --batch-size 2+```++批量命令会生成 `manifest.jsonl` 和 `failures.jsonl`。除非传入 `--overwrite`,否则已有 WAV 会+被跳过。采样和 codec codes 保存参数可通过 `python audio8_tts_infer.py --help` 查看。diff --git a/guides/models/preview-0.1b.md b/guides/models/preview-0.1b.mdnew file mode 100644index 0000000..fba2b15--- /dev/null+++ b/guides/models/preview-0.1b.md@@ -0,0 +1,37 @@+# Audio8-TTS-Preview-0.1B++**How small a zero-shot cloning TTS can be?**++[All releases](../../README.md) · [中文](preview-0.1b_zh.md) · [Model](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b) · [Demo](https://edge0-ai.github.io/Audio8_TTS/0.1B/)++A compact release for zero-shot voice cloning, using a Falcon-H1 hybrid slow AR (Mamba 2 SSM + attention). It is maintained separately from the two 0.6B releases.++## Languages and scale++- **Primary languages: Chinese and English.**+- Experimental / multilingual evaluation: German, Spanish, French, Italian, Japanese, and Korean.+- 0.1B is the release name; the current base model card reports approximately 170M main-model parameters, excluding the separate codec decoder.++See the [0.1B Model Card](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b) for architecture and Seed-TTS/CV3 results. Realistic's 23-language claim and 0.6B results do not apply to this release.++## Inference++Python 3.11+ is recommended. Follow the [installation guide](../inference.md#installation) and run from the repository root:++```bash+python audio8_tts_infer.py \+ --model Edge0/Audio8-TTS-Preview-0.1b \+ --text "Hello, this is a speech synthesis test." \+ --reference-audio reference.wav \+ --reference-text "The exact transcript of the recording." \+ --output outputs/preview-0.1b.wav+```++## Deployment++- **SGLang Omni:** Select audio8_tts_0_1b.yaml and audio8/tts-0.1b. The adapter uses independent hybrid caches with CUDA Graph disabled. [Full guide](../deployment.md).+- **ONNX INT8:** [Download](https://huggingface.co/Edge0/audio8-TTS-0.1B-ONNX-INT8) · [Dedicated runtime](../../onnx_runtime_0_1b_int8/README.md). Includes CLI, web/HTTP serving, streaming PCM, voice registration, and Windows PowerShell entry points. Do not use the 0.6B INT4 runtime with these graphs.++## License++Repository code uses [Apache License 2.0](../../LICENSE). The 0.1B base checkpoint uses [Audio8 Community License v1.0](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b#license). For separate exports such as ONNX, consult the license included in that download repository. [NOTICE](../../NOTICE).diff --git a/guides/models/preview-0.1b_zh.md b/guides/models/preview-0.1b_zh.mdnew file mode 100644index 0000000..e1f1bf7--- /dev/null+++ b/guides/models/preview-0.1b_zh.md@@ -0,0 +1,37 @@+# Audio8-TTS-Preview-0.1B++**零样本音色克隆 TTS,还能做到多小?**++[系列首页](../../README_zh.md) · [English](preview-0.1b.md) · [Model](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b) · [Demo](https://edge0-ai.github.io/Audio8_TTS/0.1B/)++面向小体积部署的零样本音色克隆型号,采用 Falcon-H1 混合 slow AR(Mamba 2 SSM + attention),与两个 0.6B 型号分别维护。++## 语言与规模++- **主要语言:中文、英语。**+- 实验 / 多语言评测:德语、西班牙语、法语、意大利语、日语、韩语。+- 发布名为 0.1B;当前基础模型卡标注主模型约 170M 参数,不含独立 codec decoder。++完整架构及 Seed-TTS/CV3 数据见 [0.1B Model Card](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b)。Realistic 的 23 语言和 0.6B 评测不适用于本版。++## 推理++推荐 Python 3.11+;先按 [安装指南](../inference_zh.md#安装)安装依赖,在仓库根目录运行:++```bash+python audio8_tts_infer.py \+ --model Edge0/Audio8-TTS-Preview-0.1b \+ --text "你好,这是一条语音合成测试。" \+ --reference-audio reference.wav \+ --reference-text "参考录音的准确文本。" \+ --output outputs/preview-0.1b.wav+```++## 部署++- **SGLang Omni:** 选择 audio8_tts_0_1b.yaml 和 audio8/tts-0.1b;适配器自动使用独立混合 cache,关闭 CUDA Graph。 [完整指南](../deployment_zh.md).+- **ONNX INT8:** [下载](https://huggingface.co/Edge0/audio8-TTS-0.1B-ONNX-INT8) · [专用运行时](../../onnx_runtime_0_1b_int8/README_zh.md). 包含命令行、网页/HTTP 服务、流式 PCM、音色注册和 Windows PowerShell 入口;不能使用 0.6B 的 INT4 运行时。++## 许可++仓库代码采用 [Apache License 2.0](../../LICENSE). 0.1B 基础权重采用 [Audio8 Community License v1.0](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.1b#license). ONNX 等独立制品以对应下载仓库附带的许可为准。 [NOTICE](../../NOTICE).diff --git a/guides/models/preview-0.6b.md b/guides/models/preview-0.6b.mdnew file mode 100644index 0000000..26755c4--- /dev/null+++ b/guides/models/preview-0.6b.md@@ -0,0 +1,124 @@+# Audio8-TTS-Preview-0.6B++[All releases](../../README.md) · [中文](preview-0.6b_zh.md) · [Model](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b) · [Demo](https://edge0-ai.github.io/Audio8_TTS/preview-0.6b/)++This page preserves the original Preview 0.6B language coverage, architecture, and evaluation. Realistic's 23-language coverage and new results belong to that release; Preview keeps its own model and usage instructions.++## Quick start++Follow the [installation instructions](../inference.md#installation), then run from the repository root:++```bash+python audio8_tts_infer.py \+ --model Edge0/Audio8-TTS-Preview-0.6b \+ --text "Welcome to Audio8 TTS Preview." \+ --output outputs/preview.wav+```++For cloning, batching, and training, continue to select `Edge0/Audio8-TTS-Preview-0.6b`.++[Inference](../inference.md) · [SGLang Omni](../deployment.md) · [ONNX INT4](../../onnx_runtime/README.md) · [SFT](../training.md)++## Supported languages++The Preview checkpoint performs best in the following languages:++| Language | Name |+|---|---|+| Cantonese | 粤语 |+| Chinese | 中文 |+| Dutch | 荷兰语 |+| English | 英语 |+| French | 法语 |+| German | 德语 |+| Italian | 意大利语 |+| Japanese | 日语 |+| Korean | 韩语 |+| Polish | 波兰语 |+| Spanish | 西班牙语 |++## Architecture++audio8_tts uses a DualAR architecture inspired by+[Fish Audio S2 Pro](https://github.com/fishaudio/fish-speech).++| Component | Configuration |+|---|---|+| Main model | 601,159,424 parameters, excluding the codec |+| Slow AR | 24 layers, width 896, 14 attention heads, 2 KV heads |+| Fast AR | 4 layers, width 896, 14 attention heads, 2 KV heads |+| Acoustic tokens | 10 codebooks, 4,096 entries per codebook |+| Codec | 44.1 kHz, 2,048 samples per model frame (~21.5 frames/s) |+| Context | Up to 2,048 packed text/audio positions |++The slow AR transformer predicts one semantic token for each audio frame. The+fast AR transformer then predicts the frame's codec codebooks, conditioned on+the slow hidden state and preceding codebooks. Static KV caches are used by+both branches during generation. The checkpoint also bundles its neural codec,+so reference encoding and waveform decoding require no separate model.++## Evaluation++Audio8 TTS Preview is the smallest model in this comparison at just **0.6B+parameters**. Despite using only a fraction of the parameters of the other+systems, it delivers results in the first tier of industry-leading SOTA TTS+models on the benchmarks below. In particular, it achieves the best English+WER and competitive Chinese CER on Seed-TTS, while remaining competitive+across the CV3 multilingual evaluation.++Lower WER/CER is better; higher SIM is better. Seed-TTS similarity values are+shown as percentages.++### Seed-TTS++| Model | Parameters | EN WER / SIM | ZH CER / SIM | Hard ZH CER / SIM |+|---|---:|---:|---:|---:|+| **Audio8 TTS Preview** | **0.6B** | **1.506** / 63.2 | 0.950 / 73.1 | 11.510 / 68.7 |+| Fish S2 Pro | 4.6B | 1.607 / 64.6 | 1.038 / 73.8 | 10.149 / 70.1 |+| Higgs Audio v2 | 4.7B | 1.524 / 66.4 | **0.806** / 72.1 | 10.622 / 69.3 |+| CosyVoice3-1.5B | 1.5B | 2.22 / 72.0 | 1.12 / 78.1 | **5.83** / **75.8** |+| MOSS-TTS | 8.5B | 1.85 / 73.4 | 1.20 / 78.8 | - |+| VoxCPM2 | 2.3B | 1.84 / **75.3** | 0.97 / **79.5** | 8.13 / 75.3 |++++### CV3 multilingual error rate++| Model | Parameters | zh | en | hard-zh | hard-en | ja | ko | de | es | fr | it | ru |+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|+| **Audio8 TTS Preview** | **0.6B** | **3.205** | **3.128** | 10.535 | 5.997 | 7.205 | 4.223 | 3.447 | 3.641 | 8.790 | 4.790 | - |+| Fish S2 Pro | 4.6B | 3.600 | 3.493 | 10.588 | 7.349 | 5.139 | **4.111** | 3.605 | 2.972 | **8.600** | 4.229 | **4.702** |+| Higgs Audio v2 | 4.7B | 3.378 | 3.404 | 10.424 | **5.754** | **4.742** | 4.260 | **3.300** | **2.929** | 9.425 | **3.555** | 5.423 |+| CosyVoice3-1.5B | 1.5B | 3.91 | 4.99 | 9.77 | 10.55 | 7.57 | 5.69 | 6.43 | 4.47 | 11.8 | 10.5 | 6.64 |+| VoxCPM2 | 2.3B | 3.65 | 5.00 | **8.55** | 8.48 | 5.96 | 5.69 | 4.77 | 3.80 | 9.85 | 4.25 | 5.21 |++++Parameter counts are calculated directly from the released weight tensors.+MOSS-TTS contains 8,489,841,664 parameters. VoxCPM2's main model contains+2,290,004,544 parameters; the separate AudioVAE is not included in the+parameter comparison.++Fish S2 Pro was reevaluated because its official evaluation uses its own+normalizer. Higgs Audio v2 was evaluated locally because concrete values were+unavailable. All other baseline values were collected from their official+reports through the [VoxCPM repository](https://github.com/OpenBMB/VoxCPM).++Different normalizers and evaluators make cross-project values reference+comparisons rather than a strictly matched ranking. Evaluation coverage does+not expand the Preview's supported-language claim beyond the 11 languages+listed above.++## Limitations and responsible use++- This is a Preview checkpoint with limited multilingual and dialect coverage.+- Very long, noisy, or inaccurate reference clips can reduce stability and+ speaker similarity.+- Generated speech can be misused for impersonation or misinformation. Obtain+ consent before cloning a voice and clearly disclose synthetic audio where+ appropriate.+- Test the model for accuracy, safety, and legal compliance before deployment.++## License++The code and these 0.6B weights use [Apache License 2.0](../../LICENSE). [NOTICE](../../NOTICE).diff --git a/guides/models/preview-0.6b_zh.md b/guides/models/preview-0.6b_zh.mdnew file mode 100644index 0000000..dcbfd8a--- /dev/null+++ b/guides/models/preview-0.6b_zh.md@@ -0,0 +1,110 @@+# Audio8-TTS-Preview-0.6B++[系列首页](../../README_zh.md) · [English](preview-0.6b.md) · [Model](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b) · [Demo](https://edge0-ai.github.io/Audio8_TTS/preview-0.6b/)++本页保留原 Preview 0.6B 的语言、架构和评测。新版 Realistic 的 23 种语言与评测仅适用于新版;Preview 的使用方式和下载入口继续保留。++## 快速开始++完成 [安装说明](../inference_zh.md#安装),然后在仓库根目录执行:++```bash+python audio8_tts_infer.py \+ --model Edge0/Audio8-TTS-Preview-0.6b \+ --text "欢迎使用 Audio8 TTS Preview。" \+ --output outputs/preview.wav+```++后续克隆、批量推理和训练继续显式使用 `Edge0/Audio8-TTS-Preview-0.6b`.++[推理](../inference_zh.md) · [SGLang Omni](../deployment_zh.md) · [ONNX INT4](../../onnx_runtime/README_zh.md) · [SFT](../training_zh.md)++## 支持语言++Preview 模型当前表现较好、推荐使用的语言如下:++| Language | 语言 |+|---|---|+| Cantonese | 粤语 |+| Chinese | 中文 |+| Dutch | 荷兰语 |+| English | 英语 |+| French | 法语 |+| German | 德语 |+| Italian | 意大利语 |+| Japanese | 日语 |+| Korean | 韩语 |+| Polish | 波兰语 |+| Spanish | 西班牙语 |++## 模型结构++audio8_tts 沿用了 [Fish Audio S2 Pro](https://github.com/fishaudio/fish-speech)+的 DualAR 架构思路。++| 组件 | 配置 |+|---|---|+| 主模型 | 601,159,424 参数,不包含 codec |+| Slow AR | 24 层、896 维、14 个 attention head、2 个 KV head |+| Fast AR | 4 层、896 维、14 个 attention head、2 个 KV head |+| 声学 token | 10 个 codebook,每个 codebook 包含 4,096 个条目 |+| Codec | 44.1 kHz,每个模型帧 2,048 个采样点,约 21.5 帧/秒 |+| 上下文 | 最多 2,048 个打包后的文本/音频位置 |++Slow AR 每个音频帧生成一个 semantic token;Fast AR 根据 slow hidden state 和当前帧已经生成的+codebook,继续生成该帧的十个 codec codebook。推理时 slow AR 和 fast AR 都使用静态 KV cache。+Checkpoint 内置神经音频 codec,不需要额外下载参考音频编码器或波形解码器。++## 评测结果++Audio8 TTS Preview 仅有 **0.6B 参数**,是本次对比中规模最小的模型。尽管参数量仅为其他模型的+一小部分,它在以下基准测试中依然达到业界领先 TTS 模型的第一梯队水平。其中,Audio8 TTS+Preview 在 Seed-TTS 上取得了最低的英文 WER 和有竞争力的中文 CER,并在 CV3 多语言评测中+保持了具有竞争力的整体表现。++WER/CER 越低越好,SIM 越高越好。Seed-TTS 的相似度统一显示为百分数。++### Seed-TTS++| 模型 | 参数量 | EN WER / SIM | ZH CER / SIM | Hard ZH CER / SIM |+|---|---:|---:|---:|---:|+| **Audio8 TTS Preview** | **0.6B** | **1.506** / 63.2 | 0.950 / 73.1 | 11.510 / 68.7 |+| Fish S2 Pro | 4.6B | 1.607 / 64.6 | 1.038 / 73.8 | 10.149 / 70.1 |+| Higgs Audio v2 | 4.7B | 1.524 / 66.4 | **0.806** / 72.1 | 10.622 / 69.3 |+| CosyVoice3-1.5B | 1.5B | 2.22 / 72.0 | 1.12 / 78.1 | **5.83** / **75.8** |+| MOSS-TTS | 8.5B | 1.85 / 73.4 | 1.20 / 78.8 | - |+| VoxCPM2 | 2.3B | 1.84 / **75.3** | 0.97 / **79.5** | 8.13 / 75.3 |++++### CV3 多语言错误率++| 模型 | 参数量 | zh | en | hard-zh | hard-en | ja | ko | de | es | fr | it | ru |+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|+| **Audio8 TTS Preview** | **0.6B** | **3.205** | **3.128** | 10.535 | 5.997 | 7.205 | 4.223 | 3.447 | 3.641 | 8.790 | 4.790 | - |+| Fish S2 Pro | 4.6B | 3.600 | 3.493 | 10.588 | 7.349 | 5.139 | **4.111** | 3.605 | 2.972 | **8.600** | 4.229 | **4.702** |+| Higgs Audio v2 | 4.7B | 3.378 | 3.404 | 10.424 | **5.754** | **4.742** | 4.260 | **3.300** | **2.929** | 9.425 | **3.555** | 5.423 |+| CosyVoice3-1.5B | 1.5B | 3.91 | 4.99 | 9.77 | 10.55 | 7.57 | 5.69 | 6.43 | 4.47 | 11.8 | 10.5 | 6.64 |+| VoxCPM2 | 2.3B | 3.65 | 5.00 | **8.55** | 8.48 | 5.96 | 5.69 | 4.77 | 3.80 | 9.85 | 4.25 | 5.21 |++++参数量直接根据已发布的权重张量统计。MOSS-TTS 包含 8,489,841,664 个参数;VoxCPM2 主模型+包含 2,290,004,544 个参数,单独的 AudioVAE 不计入本次参数量对比。++Fish S2 Pro 因官方评测使用自身 normalizer 而重新评测。Higgs Audio v2 因未公布具体数值而自行+评测。其他基线均采用 [VoxCPM 官方仓库](https://github.com/OpenBMB/VoxCPM)汇总的各项目官方值。++不同 normalizer 和评测器下的跨项目数值只能作为参考,不能视为严格同口径排名;评测覆盖也不+会将 Preview 的正式支持范围扩展到前述 11 种语言之外。++## 限制与负责任使用++- 当前是 Preview 模型,多语言和中文方言覆盖仍然有限。+- 过长、噪声较大或转写不准确的参考音频可能降低稳定性和音色相似度。+- 合成语音可能被用于冒充或传播虚假信息。克隆他人声音前应获得许可,并在适当场景明确标注合成内容。+- 部署前应针对具体业务完成准确性、安全性和合规测试。++## 许可++代码与该 0.6B 权重采用 [Apache License 2.0](../../LICENSE). [NOTICE](../../NOTICE).diff --git a/guides/models/realistic-0.6b.md b/guides/models/realistic-0.6b.mdnew file mode 100644index 0000000..8e18d04--- /dev/null+++ b/guides/models/realistic-0.6b.md@@ -0,0 +1,187 @@+# Audio8-TTS-Realistic-0.6B++[All releases](../../README.md) · [中文](realistic-0.6b_zh.md) · [Model](https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b) · [Demo](https://edge0-ai.github.io/Audio8_TTS/realistic-0.6b/)++Realistic builds on Preview with substantially improved speech realism and 23-language coverage. It retains the same compact 0.6B DualAR architecture, 44.1 kHz codec, and zero-shot voice cloning.++[](https://arxiv.org/abs/2610.14776)++Audio8 is also participating in [TTS Arena](https://tts-agi-tts-arena-v2.hf.space/), where it has received encouraging ratings from listeners. We are grateful to everyone who has listened and voted.++## Supported languages++| Coverage | Languages |+|---|---|+| 11 supported in Preview | Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Spanish |+| 12 added in Realistic | Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish, Vietnamese |++> **Note:** Language quality varies; see the per-language evaluation below. Benchmark coverage differs from the supported-language list.++## Usage and deployment++Follow the [installation instructions](../inference.md#installation), then run from the repository root:++```bash+python audio8_tts_infer.py \+ --model Edge0/Audio8-TTS-Realistic-0.6b \+ --text "Welcome to Audio8 TTS." \+ --output outputs/realistic.wav+```++[Cloning and batch inference](../inference.md) · [SGLang Omni](../deployment.md) · [SFT](../training.md)++SGLang uses the 0.6B configuration with the Realistic model selected explicitly. No Realistic ONNX export is currently available.++## Architecture++audio8_tts uses a DualAR architecture inspired by+[Fish Audio S2 Pro](https://github.com/fishaudio/fish-speech).++| Component | Configuration |+|---|---|+| Main model | 601,159,424 parameters, excluding the codec |+| Slow AR | 24 layers, width 896, 14 attention heads, 2 KV heads |+| Fast AR | 4 layers, width 896, 14 attention heads, 2 KV heads |+| Acoustic tokens | 10 codebooks, 4,096 entries per codebook |+| Codec | 44.1 kHz, 2,048 samples per model frame (~21.5 frames/s) |+| Context | Up to 2,048 packed text/audio positions |++The slow AR transformer predicts one semantic token for each audio frame. The+fast AR transformer then predicts the frame's codec codebooks, conditioned on+the slow hidden state and preceding codebooks. Static KV caches are used by+both branches during generation. The checkpoint also bundles its neural codec,+so reference encoding and waveform decoding require no separate model.++## Evaluation++Error rates are percentages (**lower is better**). Speaker similarity (SIM) is on a **0–1 scale** (**higher is better**). These metrics measure intelligibility and speaker similarity; they do not directly measure perceived speech realism.++### Benchmark summary++**Audio8-TTS-Preview** and **Audio8-TTS-Realistic**, both **0.6B**:++| Benchmark | Coverage | Preview error % ↓ | Realistic error % ↓ | Preview SIM ↑ | Realistic SIM ↑ |+|---|---|---:|---:|---:|---:|+| Seed-TTS | English / Chinese | 1.23 | 1.25 | 0.682 | 0.704 |+| CV3 | 8 shared languages | 4.80 | 4.56 | — | 0.663 |+| CV3 | 9 core languages | — | 4.86 | — | 0.661 |+| MiniMax | 24 evaluated languages | — | 5.89 | — | 0.754 |++> **Note:** Values are unweighted language macro averages. Preview results are calculated from its [published model card](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b#evaluation), with SIM converted from percentages to a 0–1 scale. The two releases were evaluated separately. A dash means the corresponding result was not reported on Preview's card.+>+> Seed-TTS averages English WER and Chinese CER, excluding hard Chinese. The shared CV3 subset is Chinese, English, Japanese, Korean, German, French, Spanish, and Italian; the 9-language set also includes Russian. CV3 averages exclude hard subsets and the 13 additional languages. MiniMax averages all 24 evaluated languages, including languages outside the supported list.++### Reference comparison++The following error rates come from the sources used in our technical report. Different evaluation pipelines and text normalization make this a reference comparison rather than a strictly matched ranking. A dash means that a result is unavailable in the comparison, not that the model lacks support.++| Model | Parameters (B) | Seed-TTS EN/ZH error ↓ | CV3 (9) error ↓ | MiniMax (24) error ↓ |+|---|---:|---:|---:|---:|+| **Audio8-TTS-Realistic** | **0.6** | **1.25** | **4.86** | **5.89** |+| Qwen3-TTS-12Hz-0.6B | 0.6 | 1.12 | — | — |+| Qwen3-TTS-1.7B | 1.7 | 1.01 | 7.73 | 27.41 |+| Higgs Audio v3 | 4 | 1.11 | 4.41 | 2.74 |+| Higgs Audio v2 | 4.7 | 1.17 | 4.49 | — |+| Fish S2 Pro | 4.6 | 1.32 | 4.49 | 4.40 |+| MOSS-TTS-v1.5 | 8.5 | 1.52 | 6.11 | 3.78 |+| CosyVoice3-1.5B | 1.5 | 1.67 | 6.89 | — |+| VoxCPM2 | 2.3 | 1.41 | 5.35 | — |++> **Sources:** Qwen Seed-TTS results are from the Qwen3-TTS report, linked from the [official Qwen3-TTS repository](https://github.com/QwenLM/Qwen3-TTS); Qwen's CV3/MiniMax results, all Higgs Audio v3 results, and MOSS-TTS-v1.5's CV3/MiniMax results are from [Boson AI's evaluation](https://huggingface.co/bosonai/higgs-audio-v3-tts-4b). Fish S2 Pro's MiniMax result is from the Fish Audio S2 report, linked from the [official Fish Speech repository](https://github.com/fishaudio/fish-speech); its Seed-TTS/CV3 results and Higgs Audio v2's results are our reevaluations. MOSS-TTS-v1.5's Seed-TTS result and the CosyVoice3/VoxCPM2 results are compiled in the [VoxCPM repository](https://github.com/OpenBMB/VoxCPM). Cross-project SIM values are omitted because speaker encoders and evaluation conventions differ.++> **Parameter conventions:** Audio8's 0.6B excludes its separate codec; Qwen and CosyVoice use release labels; Higgs Audio v3 uses its nominal **4B** release label. The remaining sizes are rounded checkpoint counts used in our report (VoxCPM2 excludes its separate AudioVAE). These counts use different conventions across systems.++### Seed-TTS detail++| Split | Metric | Preview error % ↓ | Realistic error % ↓ | Preview SIM ↑ | Realistic SIM ↑ |+|---|---|---:|---:|---:|---:|+| English | WER | 1.506 | 1.464 | 0.632 | 0.663 |+| Chinese | CER | 0.950 | 1.029 | 0.731 | 0.745 |+| Chinese, hard | CER | 11.510 | 12.797 | 0.687 | 0.702 |++> **Note:** Preview values are from its [published Seed-TTS table](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b#seed-tts), with SIM converted to a 0–1 scale.++### Multilingual detail++<details>+<summary>CV3: 22 languages and hard subsets</summary>++| Language | Metric | Error % ↓ | SIM ↑ |+|---|---|---:|---:|+| Chinese | CER | 1.814 | 0.6898 |+| English | WER | 2.907 | 0.6295 |+| Japanese | CER | 7.511 | 0.6659 |+| Korean | CER | 4.801 | 0.6962 |+| German | WER | 3.200 | 0.6669 |+| French | WER | 7.938 | 0.6355 |+| Spanish | WER | 3.389 | 0.6680 |+| Italian | WER | 4.933 | 0.6498 |+| Russian | WER | 7.258 | 0.6451 |+| Arabic | WER | 21.098 | 0.6122 |+| Finnish | WER | 32.279 | 0.7028 |+| Hindi | WER | 21.714 | 0.7245 |+| Indonesian | WER | 13.743 | 0.6682 |+| Malay | WER | 17.798 | 0.6710 |+| Dutch | WER | 10.873 | 0.6918 |+| Norwegian | WER | 17.961 | 0.6905 |+| Polish | WER | 19.233 | 0.6738 |+| Portuguese | WER | 6.871 | 0.6772 |+| Swedish | WER | 19.252 | 0.7104 |+| Thai | WER | 48.574 | 0.6691 |+| Turkish | WER | 12.567 | 0.6496 |+| Vietnamese | WER | 28.111 | 0.6790 |++Hard Chinese CER is **10.665%** and hard English WER is **4.736%**. These hard subsets are excluded from the 9-language summary. Each of the 22 regular language rows contains 500 evaluation items. Chinese, Japanese, and Korean use CER; the other regular language rows use WER.++</details>++<details>+<summary>MiniMax: all 24 evaluated languages</summary>++| Language | Metric | Error % ↓ | SIM ↑ |+|---|---|---:|---:|+| Arabic | CER | 2.403 | 0.7040 |+| Cantonese | CER | 31.349 | 0.7710 |+| Chinese | CER | 1.112 | 0.7950 |+| Czech † | WER | 8.628 | 0.7330 |+| Dutch | WER | 1.715 | 0.7050 |+| English | WER | 2.161 | 0.7740 |+| Finnish | WER | 5.072 | 0.7910 |+| French | WER | 3.198 | 0.6660 |+| German | WER | 0.898 | 0.7510 |+| Greek † | CER | 4.181 | 0.7910 |+| Hindi | CER | 9.209 | 0.8170 |+| Indonesian | WER | 2.128 | 0.7270 |+| Italian | WER | 1.958 | 0.7370 |+| Japanese | CER | 5.142 | 0.7890 |+| Korean | CER | 3.731 | 0.7850 |+| Polish | WER | 1.957 | 0.7890 |+| Portuguese | WER | 2.234 | 0.7860 |+| Romanian † | WER | 17.425 | 0.7330 |+| Russian | WER | 3.071 | 0.7490 |+| Spanish | WER | 1.350 | 0.7450 |+| Thai | CER | 6.674 | 0.7760 |+| Turkish | WER | 1.560 | 0.7650 |+| Ukrainian † | WER | 16.343 | 0.6940 |+| Vietnamese | CER | 7.966 | 0.7120 |++Each language has 100 evaluation items. Rows marked † are evaluated for completeness and are outside this release's supported-language list. The metric column preserves the evaluation suite's language-specific WER/CER choice. Cantonese is included in the 24-language average.++</details>++### Compared with Preview++Realistic expands Preview's **11-language** coverage to **23 languages**, adding **Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish, and Vietnamese**.++## Limitations and responsible use++- Quality varies by language and evaluation set. The detailed results include difficult cases, including Cantonese and some newly supported languages; the support list is not a claim of equal quality across languages.+- Noisy reference audio or mismatched transcripts can reduce stability and speaker similarity. The reference audio, transcript, and target text share a finite context budget; split long passages into manageable utterances.+- Speech may contain pronunciation errors, repetitions, omissions, or early stopping. Listen to generated audio before using it in a finished product.+- Obtain permission to clone a voice and disclose synthetic audio where appropriate.++## License and acknowledgements++The code and these 0.6B weights use [Apache License 2.0](../../LICENSE). [NOTICE](../../NOTICE).++We thank the Fish Audio team for publishing the DualAR architecture used in Fish Audio S2 Pro.diff --git a/guides/models/realistic-0.6b_zh.md b/guides/models/realistic-0.6b_zh.mdnew file mode 100644index 0000000..80519ac--- /dev/null+++ b/guides/models/realistic-0.6b_zh.md@@ -0,0 +1,185 @@+# Audio8-TTS-Realistic-0.6B++[系列首页](../../README_zh.md) · [English](realistic-0.6b.md) · [Model](https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b) · [Demo](https://edge0-ai.github.io/Audio8_TTS/realistic-0.6b/)++Realistic 在 Preview 基础上显著提升语音拟真感,并扩展到 23 种语言。保留同样的 0.6B DualAR 架构、44.1 kHz codec 和零样本音色克隆能力。++[](https://arxiv.org/abs/2610.14776)++Audio8 也参与了 [TTS Arena](https://tts-agi-tts-arena-v2.hf.space/),目前收到的听众评分令人鼓舞。感谢每一位参与试听和投票的朋友。++## 支持语言++| 范围 | 语言 |+|---|---|+| Preview 已支持的 11 种 | 粤语、中文、荷兰语、英语、法语、德语、意大利语、日语、韩语、波兰语、西班牙语 |+| Realistic 新增的 12 种 | 阿拉伯语、芬兰语、印地语、印度尼西亚语、马来语、挪威语、葡萄牙语、俄语、瑞典语、泰语、土耳其语、越南语 |++> **说明:** 语言质量有差异,详见分语言评测。评测覆盖与正式支持清单不同。++## 使用与部署++完成 [安装说明](../inference_zh.md#安装),在仓库根目录执行:++```bash+python audio8_tts_infer.py \+ --model Edge0/Audio8-TTS-Realistic-0.6b \+ --text "你好,欢迎使用 Audio8 TTS。" \+ --output outputs/realistic.wav+```++[克隆与批量推理](../inference_zh.md) · [SGLang Omni](../deployment_zh.md) · [SFT](../training_zh.md)++SGLang 使用 0.6B 配置并显式指定 Realistic 模型。当前暂无 Realistic ONNX 制品。++## 模型结构++audio8_tts 沿用了 [Fish Audio S2 Pro](https://github.com/fishaudio/fish-speech)+的 DualAR 架构思路。++| 组件 | 配置 |+|---|---|+| 主模型 | 601,159,424 参数,不包含 codec |+| Slow AR | 24 层、896 维、14 个 attention head、2 个 KV head |+| Fast AR | 4 层、896 维、14 个 attention head、2 个 KV head |+| 声学 token | 10 个 codebook,每个 codebook 包含 4,096 个条目 |+| Codec | 44.1 kHz,每个模型帧 2,048 个采样点,约 21.5 帧/秒 |+| 上下文 | 最多 2,048 个打包后的文本/音频位置 |++Slow AR 每个音频帧生成一个 semantic token;Fast AR 根据 slow hidden state 和当前帧已经生成的+codebook,继续生成该帧的十个 codec codebook。推理时 slow AR 和 fast AR 都使用静态 KV cache。+Checkpoint 内置神经音频 codec,不需要额外下载参考音频编码器或波形解码器。++## 评测++错误率使用百分数(**越低越好**);说话人相似度 SIM 使用 **0–1 范围**(**越高越好**)。这些指标衡量可懂度与音色相似度,不能直接衡量语音拟真感。++### 基准摘要++**Audio8-TTS-Preview** 与 **Audio8-TTS-Realistic**,均为 **0.6B**:++| 基准 | 覆盖 | Preview 错误率 % ↓ | Realistic 错误率 % ↓ | Preview SIM ↑ | Realistic SIM ↑ |+|---|---|---:|---:|---:|---:|+| Seed-TTS | 英语 / 中文 | 1.23 | 1.25 | 0.682 | 0.704 |+| CV3 | 8 种共有语言 | 4.80 | 4.56 | — | 0.663 |+| CV3 | 9 种核心语言 | — | 4.86 | — | 0.661 |+| MiniMax | 24 种评测语言 | — | 5.89 | — | 0.754 |++> **说明:** 数值是各语言的无权重宏平均。Preview 数据来自其[已发布 Model Card](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b#evaluation),SIM 从百分数换算为 0–1。两个版本分别评测。“—”表示 Preview 卡片未报告对应结果。+>+> Seed-TTS 平均英语 WER 和中文 CER,不含 hard Chinese。CV3 共有集为中、英、日、韩、德、法、西、意 8 种语言;9 语言集另含俄语。CV3 均值不含 hard 子集和 13 种扩展语言。MiniMax 平均全部 24 种评测语言,其中部分不在支持清单中。++### 跨模型参考比较++以下错误率来自技术报告所引用的来源。各项目评测流程及文本归一化不同,这些数据供参考,不构成严格同口径排名。“—”表示缺少比较数值,不表示该模型不支持。++| 模型 | 参数量(B) | Seed-TTS EN/ZH 错误率 ↓ | CV3(9)错误率 ↓ | MiniMax(24)错误率 ↓ |+|---|---:|---:|---:|---:|+| **Audio8-TTS-Realistic** | **0.6** | **1.25** | **4.86** | **5.89** |+| Qwen3-TTS-12Hz-0.6B | 0.6 | 1.12 | — | — |+| Qwen3-TTS-1.7B | 1.7 | 1.01 | 7.73 | 27.41 |+| Higgs Audio v3 | 4 | 1.11 | 4.41 | 2.74 |+| Higgs Audio v2 | 4.7 | 1.17 | 4.49 | — |+| Fish S2 Pro | 4.6 | 1.32 | 4.49 | 4.40 |+| MOSS-TTS-v1.5 | 8.5 | 1.52 | 6.11 | 3.78 |+| CosyVoice3-1.5B | 1.5 | 1.67 | 6.89 | — |+| VoxCPM2 | 2.3 | 1.41 | 5.35 | — |++> **来源:** Qwen 的 Seed-TTS 来自 [Qwen3-TTS 官方仓库](https://github.com/QwenLM/Qwen3-TTS)所链接的报告;其 CV3/MiniMax、Higgs Audio v3 的全部结果及 MOSS-TTS-v1.5 的 CV3/MiniMax 来自 [Boson AI 评测](https://huggingface.co/bosonai/higgs-audio-v3-tts-4b)。Fish S2 Pro 的 MiniMax 来自 [Fish Speech 官方仓库](https://github.com/fishaudio/fish-speech)所链接的报告,其 Seed-TTS/CV3 和 Higgs Audio v2 由我们重新评测。MOSS-TTS-v1.5 的 Seed-TTS 及 CosyVoice3/VoxCPM2 来自 [VoxCPM 仓库](https://github.com/OpenBMB/VoxCPM)汇总。由于说话人编码器和评测约定不同,此处不列跨项目 SIM。++> **参数口径:** Audio8 的 0.6B 不含独立 codec;Qwen 和 CosyVoice 使用发布型号;Higgs Audio v3 使用标称 **4B**。其他数值为报告中的 checkpoint 参数统计取整,VoxCPM2 不含独立 AudioVAE。各项目统计口径不同。++### Seed-TTS 详细结果++| 子集 | 指标 | Preview 错误率 % ↓ | Realistic 错误率 % ↓ | Preview SIM ↑ | Realistic SIM ↑ |+|---|---|---:|---:|---:|---:|+| English | WER | 1.506 | 1.464 | 0.632 | 0.663 |+| Chinese | CER | 0.950 | 1.029 | 0.731 | 0.745 |+| Chinese, hard | CER | 11.510 | 12.797 | 0.687 | 0.702 |++> **说明:** Preview 数值来自其[已发布的 Seed-TTS 表](https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b#seed-tts),SIM 已换算为 0–1。++### 多语言详细结果++<details>+<summary>CV3:22 种语言和 hard 子集</summary>++| 语言 | 指标 | 错误率 % ↓ | SIM ↑ |+|---|---|---:|---:|+| Chinese | CER | 1.814 | 0.6898 |+| English | WER | 2.907 | 0.6295 |+| Japanese | CER | 7.511 | 0.6659 |+| Korean | CER | 4.801 | 0.6962 |+| German | WER | 3.200 | 0.6669 |+| French | WER | 7.938 | 0.6355 |+| Spanish | WER | 3.389 | 0.6680 |+| Italian | WER | 4.933 | 0.6498 |+| Russian | WER | 7.258 | 0.6451 |+| Arabic | WER | 21.098 | 0.6122 |+| Finnish | WER | 32.279 | 0.7028 |+| Hindi | WER | 21.714 | 0.7245 |+| Indonesian | WER | 13.743 | 0.6682 |+| Malay | WER | 17.798 | 0.6710 |+| Dutch | WER | 10.873 | 0.6918 |+| Norwegian | WER | 17.961 | 0.6905 |+| Polish | WER | 19.233 | 0.6738 |+| Portuguese | WER | 6.871 | 0.6772 |+| Swedish | WER | 19.252 | 0.7104 |+| Thai | WER | 48.574 | 0.6691 |+| Turkish | WER | 12.567 | 0.6496 |+| Vietnamese | WER | 28.111 | 0.6790 |++Hard Chinese CER 为 **10.665%**,hard English WER 为 **4.736%**;两个 hard 子集均不计入 9 语言摘要。22 种常规语言各有 500 条评测;中文、日语和韩语使用 CER,其余使用 WER。++</details>++<details>+<summary>MiniMax:全部 24 种评测语言</summary>++| 语言 | 指标 | 错误率 % ↓ | SIM ↑ |+|---|---|---:|---:|+| Arabic | CER | 2.403 | 0.7040 |+| Cantonese | CER | 31.349 | 0.7710 |+| Chinese | CER | 1.112 | 0.7950 |+| Czech † | WER | 8.628 | 0.7330 |+| Dutch | WER | 1.715 | 0.7050 |+| English | WER | 2.161 | 0.7740 |+| Finnish | WER | 5.072 | 0.7910 |+| French | WER | 3.198 | 0.6660 |+| German | WER | 0.898 | 0.7510 |+| Greek † | CER | 4.181 | 0.7910 |+| Hindi | CER | 9.209 | 0.8170 |+| Indonesian | WER | 2.128 | 0.7270 |+| Italian | WER | 1.958 | 0.7370 |+| Japanese | CER | 5.142 | 0.7890 |+| Korean | CER | 3.731 | 0.7850 |+| Polish | WER | 1.957 | 0.7890 |+| Portuguese | WER | 2.234 | 0.7860 |+| Romanian † | WER | 17.425 | 0.7330 |+| Russian | WER | 3.071 | 0.7490 |+| Spanish | WER | 1.350 | 0.7450 |+| Thai | CER | 6.674 | 0.7760 |+| Turkish | WER | 1.560 | 0.7650 |+| Ukrainian † | WER | 16.343 | 0.6940 |+| Vietnamese | CER | 7.966 | 0.7120 |++每种语言有 100 条评测。† 标记的语言用于完整评测,不在本版正式支持清单中。指标列保留该评测集对各语言的 WER/CER 选择。粤语计入 24 语言均值。++</details>++### 相较 Preview++Realistic 将 Preview 的 **11 种语言**扩展到 **23 种**,新增**阿拉伯语、芬兰语、印地语、印度尼西亚语、马来语、挪威语、葡萄牙语、俄语、瑞典语、泰语、土耳其语和越南语**。++## 限制与负责任使用++- 质量随语言和评测集而变化,支持清单不代表各语言表现相同;分语言结果保留粤语及新增语言中的困难情况。+- 参考音频噪声或转写不匹配可能降低稳定性和音色相似度。参考音频、转写和目标文本共用有限上下文,长文本应拆分。+- 合成可能出现读音错误、重复、遗漏或提前结束,用于成品前请试听。+- 克隆他人音色前应获得许可,并在适当场景标明合成内容。++## 许可与致谢++代码与本版 0.6B 权重采用 [Apache License 2.0](../../LICENSE). [NOTICE](../../NOTICE).++感谢 Fish Audio 团队公开 Fish Audio S2 Pro 使用的 DualAR 架构。diff --git a/guides/training.md b/guides/training.mdnew file mode 100644index 0000000..e0b4559--- /dev/null+++ b/guides/training.md@@ -0,0 +1,72 @@+# Supervised fine-tuning (SFT)++[All releases](../README.md) · [中文](training_zh.md) · [Inference](inference.md)++These instructions retain the 0.6B SFT workflow. Run from the repository root and explicitly choose the release to fine-tune.++```bash+export MODEL=Edge0/Audio8-TTS-Realistic-0.6b+# Preview 0.6B: export MODEL=Edge0/Audio8-TTS-Preview-0.6b+```++Install the training dependencies first:++```bash+pip install -r requirements-train.txt+```++### 1. Create a raw manifest++The target `audio` field is required. `reference_audio` and `reference_text`+are optional, but must be provided together.++```json+{"id":"utt_001","text":"Target transcript","audio":"audio/target.wav","reference_audio":"audio/reference.wav","reference_text":"Reference transcript"}+{"id":"utt_002","text":"Another transcript","audio":"audio/another.wav"}+```++### 2. Precompute codec indices++```bash+python audio8_tts_prepare.py \+ --model "${MODEL}" \+ --input-jsonl data/train.jsonl \+ --output-jsonl prepared_data/train.jsonl \+ --batch-size 4+```++The prepared manifest points to validated `[10, T]` NumPy arrays using paths+relative to the prepared manifest. Existing valid arrays are reused unless+`--overwrite` is passed.++### 3. Train++Single GPU:++```bash+MODEL="${MODEL}" \+TRAIN_JSONL=prepared_data/train.jsonl \+NPROC_PER_NODE=1 \+bash audio8_tts_sft.sh+```++Eight GPUs on one node:++```bash+MODEL="${MODEL}" \+TRAIN_JSONL=prepared_data/train.jsonl \+NPROC_PER_NODE=8 \+BATCH_SIZE=2 \+GRADIENT_ACCUMULATION_STEPS=8 \+bash audio8_tts_sft.sh+```++For multi-node training, set `NNODES`, `NODE_RANK`, `MASTER_ADDR`, and+`MASTER_PORT` on each node. Common hyperparameters and output paths can be+overridden through the environment variables in `audio8_tts_sft.sh`; additional+Transformers arguments may be appended to the command.++SFT optimizes both the slow semantic/EOS objective and the fast codebook+teacher-forcing objective. Set `FREEZE_SLOW_AR=true` or `FREEZE_FAST_AR=true`+when adapting only one branch. The exported directory remains loadable with+standard `AutoModel` and `AutoProcessor` APIs using `trust_remote_code=True`.diff --git a/guides/training_zh.md b/guides/training_zh.mdnew file mode 100644index 0000000..a676553--- /dev/null+++ b/guides/training_zh.md@@ -0,0 +1,68 @@+# 监督微调(SFT)++[系列首页](../README_zh.md) · [English](training.md) · [推理](inference_zh.md)++以下保留 0.6B SFT 流程。从仓库根目录执行,并显式选择需要微调的版本。++```bash+export MODEL=Edge0/Audio8-TTS-Realistic-0.6b+# Preview 0.6B: export MODEL=Edge0/Audio8-TTS-Preview-0.6b+```++先安装训练依赖:++```bash+pip install -r requirements-train.txt+```++### 1. 创建原始数据 manifest++目标 `audio` 必填;`reference_audio` 和 `reference_text` 可选,但必须同时出现。++```json+{"id":"utt_001","text":"目标音频文本","audio":"audio/target.wav","reference_audio":"audio/reference.wav","reference_text":"参考音频文本"}+{"id":"utt_002","text":"另一条文本","audio":"audio/another.wav"}+```++### 2. 预计算 codec indices++```bash+python audio8_tts_prepare.py \+ --model "${MODEL}" \+ --input-jsonl data/train.jsonl \+ --output-jsonl prepared_data/train.jsonl \+ --batch-size 4+```++生成的 manifest 使用相对路径指向经过校验的 `[10, T]` NumPy 数组。除非传入+`--overwrite`,否则已有的有效数组会被复用。++### 3. 开始训练++单卡:++```bash+MODEL="${MODEL}" \+TRAIN_JSONL=prepared_data/train.jsonl \+NPROC_PER_NODE=1 \+bash audio8_tts_sft.sh+```++单机八卡:++```bash+MODEL="${MODEL}" \+TRAIN_JSONL=prepared_data/train.jsonl \+NPROC_PER_NODE=8 \+BATCH_SIZE=2 \+GRADIENT_ACCUMULATION_STEPS=8 \+bash audio8_tts_sft.sh+```++多机训练时,在每个节点设置 `NNODES`、`NODE_RANK`、`MASTER_ADDR` 和 `MASTER_PORT`。常用超参+和输出路径可通过 `audio8_tts_sft.sh` 中列出的环境变量覆盖,也可以在命令末尾附加 Transformers+参数。++SFT 同时优化 slow semantic/EOS loss 和 fast codebook teacher-forcing loss。只训练其中一个分支时,+可设置 `FREEZE_SLOW_AR=true` 或 `FREEZE_FAST_AR=true`。导出目录仍可通过标准 `AutoModel` 和+`AutoProcessor` 接口配合 `trust_remote_code=True` 加载。