mlx-community/Fun-CosyVoice3-0.5B-2512-4bit
This model was converted to MLX format from FunAudioLLM/Fun-CosyVoice3-0.5B-2512 using mlx-audio-plus version 0.1.4.
This model uses 4-bit quantization for the Qwen2 LLM backbone, reducing memory usage while maintaining quality.
Usage
pip install -U mlx-audio-plus
Inference Modes
| Mode | Parameters | Description |
|---|---|---|
| Cross-lingual | ref_audio | Zero-shot TTS (default) |
| Zero-shot | ref_audio + ref_text | Better quality with transcription |
| Instruct | ref_audio + instruct_text | Style control (e.g., "speak slowly") |
| Voice Conversion | source_audio + ref_audio | Convert audio to target voice |
Command line
# Cross-lingual (default)
mlx_audio.tts --model mlx-community/Fun-CosyVoice3-0.5B-2512-4bit --text "Hello!" --ref_audio ref.wav
# Zero-shot (with transcription)
mlx_audio.tts --model mlx-community/Fun-CosyVoice3-0.5B-2512-4bit --text "Hello!" --ref_audio ref.wav --ref_text "Transcription of ref audio."
# Instruct (style control)
mlx_audio.tts --model mlx-community/Fun-CosyVoice3-0.5B-2512-4bit --text "Hello!" --ref_audio ref.wav --instruct_text "Speak slowly and calmly"
# Voice Conversion
mlx_audio.tts --model mlx-community/Fun-CosyVoice3-0.5B-2512-4bit --source_audio source.wav --ref_audio ref.wav
Python
from mlx_audio.tts.generate import generate_audio
generate_audio(
text="Hello, this is CosyVoice 3 on MLX!",
model="mlx-community/Fun-CosyVoice3-0.5B-2512-4bit",
ref_audio="reference.wav",
file_prefix="output",
)