Part of the OmniVoice MLX collection.

OmniVoice-fp32 (MLX)

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 fp32
Best for highest precision 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

hf download mlx-community/OmniVoice-fp32 --local-dir ./models/OmniVoice-fp32

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:

.venv/bin/python scripts/infer_mlx.py \
  --model ./models/OmniVoice-fp32 \
  --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

OmniVoice-fp32/
β”œβ”€β”€ config.json
β”œβ”€β”€ model.safetensors / shards
β”œβ”€β”€ tokenizer files
β”œβ”€β”€ 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

@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},
}
Downloads last month
49
Safetensors
Model size
0.6B params
Tensor type
F32
Β·
I64
Β·
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for mlx-community/OmniVoice-fp32

Finetuned
Qwen/Qwen3-0.6B
Finetuned
k2-fsa/OmniVoice
Finetuned
(86)
this model

Space using mlx-community/OmniVoice-fp32 1

Collection including mlx-community/OmniVoice-fp32

Paper for mlx-community/OmniVoice-fp32