Skip to content

SOTA-Class TTS at Compact Scale

Python
2.1 KB
Raw Blame

Audio8-TTS-Preview-0.1B

How small a zero-shot cloning TTS can be?

All releases · 中文 · Model · Demo

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 for architecture and Seed-TTS/CV3 results. Realistic Air's 23-language claim and 0.6B results do not apply to this release.

Inference

Python 3.11+ is recommended. Follow the installation guide and run from the repository root:

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.
  • ONNX INT8: Download · Dedicated runtime. 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. The 0.1B base checkpoint uses Audio8 Community License v1.0. For separate exports such as ONNX, consult the license included in that download repository. NOTICE.