Apple MLX weights for
k2-fsa/OmniVoice, a massively multilingual zero-shot TTS model. This is a community MLX conversion for Apple Silicon; the upstream model card, license, and repository remain authoritative for non-MLX usage.
1hf download mlx-community/OmniVoice-4bit --local-dir ./models/OmniVoice-4bit
2
3git clone https://github.com/ailuntx/OmniVoice-MLX.git
4cd OmniVoice-MLX
5python -m venv .venv
6.venv/bin/pip install -e .
1.venv/bin/python scripts/infer_mlx.py \
2 --model ./models/OmniVoice-4bit \
3 --text "Hello from OmniVoice MLX." \
4 --language en \
5 --output output.wav
1OmniVoice-4bit/
2├── config.json
3├── model.safetensors / shards
4├── tokenizer files
5├── audio_tokenizer/
6└── mlx_manifest.json
Local MLX smoke tests were used during conversion. For voice cloning checks, use a full audio tokenizer; slim tokenizer assets can decode audio but do not provide a reliable speaker-encoding path.
License follows the upstream OmniVoice release.
1@misc{omnivoice-mlx,
2 title = {OmniVoice-MLX: Apple MLX port of OmniVoice},
3 author = {ailuntx},
4 year = {2026},
5 url = {https://github.com/ailuntx/OmniVoice-MLX},
6}
7
8@article{zhu2026omnivoice,
9 title = {OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models},
10 author = {Zhu, Han and Ye, Lingxuan and Kang, Wei and Yao, Zengwei and Guo, Liyong and Kuang, Fangjun and Han, Zhifeng and Zhuang, Weiji and Lin, Long and Povey, Daniel},
11 journal = {arXiv preprint arXiv:2604.00688},
12 year = {2026},
13}
