vox-cosyvoice 0.1.2

CosyVoice 2 TTS adapter for Vox

vox-cosyvoice

vox-cosyvoice provides a Vox TTS adapter for CosyVoice 2.

Adapters:

• cosyvoice2-tts-torch - CosyVoice 2 zero-shot streaming backend

Install

pip install vox-cosyvoice

Runtime Dependencies

The adapter package is intentionally light. The official CosyVoice source is checked out on demand from GitHub into $VOX_HOME/runtime/cosyvoice/CosyVoice, and CosyVoice-specific Python dependencies are installed into the isolated target runtime $VOX_HOME/runtime/cosyvoice.

The shared Vox GPU stack remains owned by the base Vox environment; the adapter does not install its own Torch/CUDA/server runtime packages.

CosyVoice imports whisper.log_mel_spectrogram and whisper.tokenizer.Tokenizer for frontend features. Vox provides those narrow compatibility surfaces inside the isolated runtime instead of installing the full openai-whisper package and its duplicate GPU stack. Matcha-TTS also imports matplotlib through training/plotting utilities during module import; Vox provides an inference-only compatibility package instead of installing the full plotting stack.

Use with Vox

vox pull cosyvoice2-tts-torch:0.5b
vox run cosyvoice2-tts-torch:0.5b "Hello from CosyVoice 2"

CosyVoice 2 is best used with a reference voice. Pass reference_audio and reference_text through the Vox API, or use a voice value that points to a local WAV file.