VoxCPM2 - 4-bit quantized#
MLX port of openbmb/VoxCPM2 — a 2B-parameter multilingual TTS model with 48kHz studio-quality output, voice cloning, and voice design.
4-bit quantized (LM layers only, VAE/DiT at full precision). Fastest, smallest, with minimal quality loss.
Features#
30 languages — including English, Chinese, Indonesian, Japanese, Korean, and more48kHz output — studio-quality audioVoice Design — create voices from text descriptions (no reference audio needed)Voice Cloning — clone any voice from a short audio reference4 generation modes — zero-shot, continuation, reference cloning, combinedUsage#
pip install mlx-audio
# Zero-shot
python -m mlx_audio.tts.generate --model mlx-community/VoxCPM2-4bit --text "Hello world" --verbose
# Voice design
python -m mlx_audio.tts.generate --model mlx-community/VoxCPM2-4bit \
--text "Hello world" \
--instruct "A young woman, gentle and sweet voice"
# Voice cloning
python -m mlx_audio.tts.generate --model mlx-community/VoxCPM2-4bit \
--text "Hello world" \
--ref_audio speaker.wav --ref_text "reference text"Python API#
from mlx_audio.tts import load_model
model = load_model("mlx-community/VoxCPM2-4bit")
# Generate
for result in model.generate(
text="Hello, this is VoxCPM2 on Apple Silicon.",
inference_timesteps=7,
cfg_value=2.0,
):
print(f"Duration: {result.audio_duration}")Variant Size RTF (7 timesteps) bf16 4.96 GB 0.48x 8-bit 3.23 GB 0.85x 4-bit 2.30 GB 0.90x
RTF = Real-Time Factor (>1.0 = faster than realtime)
Original Model#
Converted with mlx-audio .
mlx-community <huggingface.co/mlx-community>Newsgroups: alt.freeweights.speechSubject: 
onnx-community