phoonnx-chatterbox

Mirror of the base (English) Chatterbox model by Resemble AI, exported to ONNX for phoonnx, OpenVoiceOS's offline TTS library. Chatterbox is an autoregressive codec-LM: a Llama-based language model generates discrete speech tokens conditioned on text and a reference speaker, which are then decoded to a waveform. It is phoonnx's first autoregressive engine, and its standout feature is zero-shot cloning from a reference clip with no transcription needed, plus an exaggeration control for expressiveness. See docs/training/engines/chatterbox.md for the full architecture and docs/cloning.md for the cloning walkthrough.

The ONNX export (onnx-community/chatterbox-ONNX) splits the model into four graphs, all present here: embed_tokens, language_model_q4 (4-bit quantized Llama LM), speech_encoder, and conditional_decoder, plus the matching tokenizer.json.

Related repositories: phoonnx-chatterbox-turbo (faster GPT-2-based variant), phoonnx-chatterbox-multilingual (30 languages), phoonnx-chatterbox-lahgtna (Arabic dialects).

Usage

Cloning support requires pip install phoonnx[cloning].

import wave
from phoonnx.model_manager import TTSModelManager

manager = TTSModelManager()
manager.load()
manager.merge_default_voices()

voice = manager.voices["chatterbox/base/en"].load()
with wave.open("out.wav", "wb") as wav_file:
    voice.synthesize_wav(
        "Hello, this is a test.",
        wav_file,
        speaker_reference="reference.wav",  # optional: clone this speaker
    )

License

Resemble AI released Chatterbox, including this base checkpoint, under the MIT license (see ResembleAI/chatterbox and the onnx-community/chatterbox-ONNX export). phoonnx itself is Apache-2.0.