A research project focused on developing a lightweight neural voice cloning and speech synthesis system, optimized for deployment on edge devices.
| Name | Last commit message | Last commit date |
|---|---|---|
| .github | ||
| docs | ||
| reverb_tts | ||
| tests | ||
| .gitignore 146 B | ||
| .pre-commit-config.yaml 225 B | ||
| LICENSE 11.1 KB | ||
| pyproject.toml 1.4 KB | ||
| README.md 7.5 KB |
Reverb-TTS
Reverb is a text-to-speech model designed for high-quality voice cloning: give it a few seconds of someone speaking and the transcript, and it speaks any text in that voice. This repository holds the Python package that runs it.
- Zero-shot cloning from 3 to 10 seconds of speech and its transcript, with no fine-tuning.
- 7 languages: English, German, French, Spanish, Italian, Polish, and Portuguese. A voice can speak any of them.
- 0.96% word error rate on the Open TTS Leaderboard's MiniMax test set, lower than every model on the leaderboard.
- 297M parameters, plus a 22M-parameter codec decoder.
- 232.3× real time on an NVIDIA H200 and 164.9× on an RTX 5090 at batch 32, over 20× on an Apple M4 Pro and an iPhone 17 at batch 1.
- Weights: fp32, fp16, int8, and int4 safetensors, and ExecuTorch programs for MLX, Vulkan, and QNN.
Install
pip install reverb-tts
Python 3.10 or later. The weights download from Hugging Face on first use.
Quick start
import soundfile as sf
from reverb_tts import TTS
tts = TTS.from_pretrained()
voice = tts.voice("reference.wav", "The exact transcript of the reference recording.", language="en")
audio = tts.generate("Hello! This is my cloned voice.", voice, language="en")
sf.write("output.wav", audio, tts.sample_rate)
No recording at hand? Use a preset voice:
voice = tts.preset_voice("pl_male_standard") # tts.preset_voices() lists them all
audio = tts.generate("Dzień dobry! Mówię po polsku.", voice, language="pl")
TTS.from_pretrained picks CUDA, Apple's MPS, or the CPU, in that order. Long texts are split at sentence ends.
Text preprocessing
Reverb reads text in spoken form: it was trained on text with numbers, dates, money, units, symbols, and abbreviations written out as words. Normalizing the text is therefore a required step before generating, not an optional extra.
- With the Python package, it's done for you:
generateandvoicerun Verbalis on the text and the transcript by default, so you can pass text as you'd write it ("$49" is read as "forty-nine dollars"). - Anywhere else (the ExecuTorch programs, ruNNtime, or
prepare=False), the text reaches the model exactly as you give it. Normalize it yourself first. We strongly recommend Verbalis, the normalizer the package uses, with C++, TypeScript, and Python APIs for all seven of Reverb's languages.
In the browser
ruNNtime runs Reverb on the GPU in the browser through
WebGPU. Install it with npm install runntime and follow its text-to-speech guide.
Documentation
- Overview: what the model is, how it was made, and how it scores.
- Usage: every use case, from the quick start to sampler settings.
- Reference: the safetensors variants, the ExecuTorch programs, and the repositories' layout.
Benchmarks
On the Open TTS Leaderboard's MiniMax test set, across seven languages, all with voice cloning: WER, speaker similarity (SIM-o), and speed on an NVIDIA H200 at batch 32.
| Model | Parameters | WER ↓ | SIM-o ↑ | H200, × real time ↑ |
|---|---|---|---|---|
| Reverb | 319M | 1.03% | 0.817 | 125.2× |
| OmniVoice | 634M | 1.13% | 0.833 | 27.1× |
| VoxCPM2 | 2,336M | 1.20% | 0.812 | 2.0× |
| Fish S2 Pro | 4.95B | 1.35% | 0.761 | 0.40× |
| MOSS-TTS | 8.49B | 1.84% | 0.774 | 19.4× |
| Chatterbox | 0.80B | 2.76% | 0.715 | 2.0× |
| Higgs TTS 3 | 4.65B | 6.59% | 0.754 | 72.9× |
The full tables, the Apple M4 Pro and iPhone 17 speeds, and the method are in the overview.
Use it responsibly
Clone only voices you have the right to use. Get the speaker's consent, don't impersonate real people, and don't use generated speech to deceive, defraud, or harass. Tell your listeners when they hear a synthetic voice.
License and acknowledgments
- Code: Apache 2.0; see LICENSE.
- Model weights: CC BY-NC 4.0, for non-commercial use with attribution.
Reverb's weights were initialized from OmniVoice by the k2-fsa team, and its target-length rule and prompt format come from OmniVoice's code. We thank its authors for releasing the model and code openly. Reverb uses the Higgs Audio v2 tokenizer as its audio codec, and Verbalis prepares its text.
Community Discord
Join the Software Mansion Community Discord to chat about Reverb or other Software Mansion libraries.
Reverb is created by Software Mansion AI Labs
We're an end-to-end AI development lab focused on local and hybrid AI. We build and fine-tune models, deploy them anywhere and turn them into agents that solve real business problems. Hire us.