---
license: cc-by-nc-4.0
base_model: [k2-fsa/OmniVoice]
pipeline_tag: text-to-speech
library_name: onnxruntime
tags: [text-to-speech, voice-cloning, zero-shot, multilingual, onnx]
---
# OmniVoice — corrected bidirectional ONNX export (int4 variant)
Self-contained **int4** variant (smallest download: 4-bit llm_decoder weights, dequantized to fp32 at compute → clean on CUDA; audio_embeddings is bf16, lossless). Backbone at repo root + `audio_tokenizer/` (Higgs fp32) + `voices/`. Siblings: [-bf16](https://huggingface.co/jiangzhuo9357/omnivoice-onnx-bidi-bf16) (default), [-fp32](https://huggingface.co/jiangzhuo9357/omnivoice-onnx-bidi-fp32).
Bidirectional re-export of [k2-fsa/OmniVoice](https://huggingface.co/k2-fsa/OmniVoice) for [Sokuji](https://github.com/kizuna-ai-lab/sokuji) (#351); plain onnxruntime.
## License — NON-COMMERCIAL
**CC-BY-NC-4.0**, inherited from k2-fsa/OmniVoice (Emilia training set). Unofficial re-export, not affiliated with/endorsed by the authors. Some re-uploads mislabel it apache-2.0 — incorrect for the weights; do not use commercially.