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.

TL;DR


Variant	default fp32 entry
Best for	default entry; same role as the fp32 baseline
Runtime	`ailuntx/OmniVoice-MLX`
Official code	`k2-fsa/OmniVoice`
Format	MLX safetensors + tokenizer/config assets
Hardware	Apple Silicon recommended; HF Spaces Linux CPU fallback can be slow

Quick Start

bash

hf download mlx-community/OmniVoice --local-dir ./models/OmniVoice

git clone https://github.com/ailuntx/OmniVoice-MLX.git
cd OmniVoice-MLX
python -m venv .venv
.venv/bin/pip install -e .

Run a minimal generation from the MLX helper repository:

bash

.venv/bin/python scripts/infer_mlx.py \
  --model ./models/OmniVoice \
  --text "Hello from OmniVoice MLX." \
  --language en \
  --output output.wav

Variants

Variant	Best for
`OmniVoice`	default entry
`OmniVoice-fp32`	high-precision baseline
`OmniVoice-bfloat16`	high-quality Apple Silicon use
`OmniVoice-8bit`	smaller local checkpoint
`OmniVoice-4bit`	smallest checkpoint and Space default

Layout

text

OmniVoice/
├── config.json
├── model.safetensors / shards
├── tokenizer files
├── audio_tokenizer/
└── mlx_manifest.json

Conversion Notes

Component	Source	MLX handling
main model	`k2-fsa/OmniVoice`	converted to MLX weights
tokenizer/config	official checkpoint	copied for runtime compatibility
audio tokenizer	official OmniVoice assets	included as a required subcomponent

Validation

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

License follows the upstream OmniVoice release.

Citation

bibtex

@misc{omnivoice-mlx,
  title  = {OmniVoice-MLX: Apple MLX port of OmniVoice},
  author = {ailuntx},
  year   = {2026},
  url    = {https://github.com/ailuntx/OmniVoice-MLX},
}

@article{zhu2026omnivoice,
  title   = {OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models},
  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},
  journal = {arXiv preprint arXiv:2604.00688},
  year    = {2026},
}