# OmniVoice — card for AI agents

Zero-shot TTS covering 600+ languages, with voice cloning from a short clip and voice design by speaker attributes.

Category: Voiceover (text-to-speech) · maker k2-fsa · checked 2026-11-10

## Price
- Price: Free
- Free tier: Open weights, free to download.

## License
Code Apache-2.0; pre-trained model CC-BY-NC (model card); audio tokenizer under the Boson Higgs Audio 2 Community License — https://huggingface.co/k2-fsa/OmniVoice/blob/main/audio_tokenizer/LICENSE

- commercial: **no** (Non-commercial only)
  > The pre-trained model is licensed under the CC-BY-NC due to constraints from its training data (e.g., Emilia).
  (https://huggingface.co/k2-fsa/OmniVoice)
- attribution: **required** (Credit required)
  > prominently display “Built with Higgs Materials licensed from Boson AI USA, Inc., Copyright Boson AI USA, Inc., All Rights Reserved
  (https://huggingface.co/k2-fsa/OmniVoice/blob/main/audio_tokenizer/LICENSE)
- in app: **no** (Cannot ship inside an app)
  > The pre-trained model is licensed under the CC-BY-NC due to constraints from its training data (e.g., Emilia).
  (https://huggingface.co/k2-fsa/OmniVoice)
- youtube: **no** (Not for monetized YouTube)
  > The pre-trained model is licensed under the CC-BY-NC due to constraints from its training data (e.g., Emilia).
  (https://huggingface.co/k2-fsa/OmniVoice)

## Rules to follow
- The code is Apache-2.0; the audio tokenizer shipped with the weights is under the Boson Higgs Audio 2 Community License (Llama 3 based).
- Products or services above 100,000 annual active users need an expanded license from Boson AI.
- A product or service using it must show "Built with Higgs Materials licensed from Boson AI" and follow the Llama 3 acceptable-use policy.
- Outputs may not be used to improve other large language models.
- The pre-trained OmniVoice model is CC-BY-NC (non-commercial), per its model card.

## Can an AI agent use it by itself?
**No** · package: Agent installs a package or loads it from a CDN

## How to connect
- Package (npm/PyPI)
- PyPI: `pip install omnivoice`
- Repository: https://github.com/k2-fsa/OmniVoice
- Website: https://huggingface.co/k2-fsa/OmniVoice

## Notes
Hugging Face API, Oct 25, 2026: 2184462 downloads, 1592 likes; GitHub k2-fsa/OmniVoice: 15860 stars.

## Numbers (AgentGid data)
- Gid Score: 43
- GitHub stars: 16,699
- Downloads last week: 43,074

## Lines to paste into your prompt or AGENTS.md

```
- OmniVoice (Voiceover (text-to-speech)). Zero-shot TTS covering 600+ languages, with voice cloning from a short clip and voice design by speaker attributes. Price: free; non-commercial only; credit required; cannot ship inside an app.
  Connect: PyPI omnivoice.
  Rule: The code is Apache-2.0; the audio tokenizer shipped with the weights is under the Boson Higgs Audio 2 Community License (Llama 3 based).
  Rule: Products or services above 100,000 annual active users need an expanded license from Boson AI.
  Rule: A product or service using it must show "Built with Higgs Materials licensed from Boson AI" and follow the Llama 3 acceptable-use policy.
  Rule: Outputs may not be used to improve other large language models.
  Rule: The pre-trained OmniVoice model is CC-BY-NC (non-commercial), per its model card.
  License: https://huggingface.co/k2-fsa/OmniVoice/blob/main/audio_tokenizer/LICENSE (checked 2026-11-10; details: https://agentgid.com/assets/omnivoice/)
```

## Links
- AgentGid page: https://agentgid.com/assets/omnivoice/
- Is it free for commercial use? https://agentgid.com/assets/omnivoice/commercial-use/
- Source: https://huggingface.co/k2-fsa/OmniVoice
- Source: https://huggingface.co/k2-fsa/OmniVoice/blob/main/audio_tokenizer/LICENSE
- Source: https://github.com/k2-fsa/OmniVoice
- Source: https://github.com/k2-fsa/OmniVoice/blob/master/LICENSE
- Source: https://pypi.org/project/omnivoice/


## OmniVoice: speech in 600+ languages, non-commercial

OmniVoice is a free AI model that reads text aloud in more than 600 languages. It can copy a voice from a few seconds of recorded speech, or make a new voice from a short description like "female, low pitch". The program code is open, but the model itself may only be used for non-commercial work.

### Where to use it
- Speech for rare languages in research: Make spoken audio for a language that common voice tools skip, for a study, a school project or a language archive.
- Your own voice from a short clip: Record 3 to 10 seconds of your voice, add what you said, and the model reads new text in that voice. Copying someone else's voice without permission is forbidden.
- A new voice from a description: No recording needed. Describe the voice by gender, age, pitch, accent or a whisper, and the model makes one.
- Laughs, sighs and exact pronunciation: Put a laugh or a sigh into the text, or spell out how a word should sound, so lines in a story or game feel natural.
- Many audio lines in one run: Turn a long list of lines into audio files at once, for example all phrases of a language course, spread over several graphics cards.
- Trying it without code: Start a local web page where you type text, upload a voice sample and listen, before deciding to build anything.

### How to start
- Install PyTorch for your machine first (the README shows the commands for NVIDIA, Apple Silicon and Intel), then install OmniVoice from PyPI.

```
pip install torch==2.8.0 torchaudio==2.8.0
pip install omnivoice
```

- Generate speech from the command line; the model downloads from Hugging Face on first use.

```
omnivoice-infer --model k2-fsa/OmniVoice \
    --text "This is a test for text to speech." \
    --instruct "male, British accent" \
    --output hello.wav
```


### License
Non-commercial only. The code is Apache-2.0, but the model is CC-BY-NC because of its training data, and a bundled part adds Boson AI terms with a credit line and a 100,000-user cap.

### Good to know
- The model may not be used commercially. Only the program code is under the open Apache license.
- Copying a voice without permission, impersonation and scams are strictly forbidden by the authors.
- Voice design from a description was trained only on Chinese and English. In other languages it may sound unstable.
- Keep the voice sample short, 3 to 10 seconds; longer clips make it slower and can make the copy worse.
- Products that use the bundled audio part must show a "Built with Higgs Materials" notice from Boson AI.

Checked 2026-11-10 against: https://huggingface.co/k2-fsa/OmniVoice, https://huggingface.co/k2-fsa/OmniVoice/raw/main/README.md, https://huggingface.co/k2-fsa/OmniVoice/raw/main/audio_tokenizer/LICENSE, https://huggingface.co/api/models/k2-fsa/OmniVoice, https://github.com/k2-fsa/OmniVoice, https://github.com/k2-fsa/OmniVoice/blob/master/README.md, https://github.com/k2-fsa/OmniVoice/blob/master/LICENSE, https://pypi.org/pypi/omnivoice/json

### Mistakes to avoid
- Don't: Adding OmniVoice to a paid product because the package says Apache-2.0 → Do: Treat the model as non-commercial; for a product, choose a model whose weights allow commercial use (The Apache license covers the code only; the model card puts the weights under CC-BY-NC.)
- Don't: device_map="cuda:0" on a Mac → Do: device_map="mps" on Apple Silicon (or "xpu" on Intel Arc) (The README example targets NVIDIA cards.)
- Don't: A 30-second voice sample for cloning → Do: Trim the sample to 3–10 seconds, in the same language as the target speech (Longer samples slow it down and can lower quality; a sample in another language adds an accent.)
- Don't: Setting HF_ENDPOINT="https://hf-mirror.com" or uv sync with the aliyun mirror by default → Do: Download from huggingface.co and PyPI unless the person chose a mirror (The README offers third-party mirrors only as a workaround for connection trouble.)

### Questions
- Can I use OmniVoice in a paid app or for client work? No. The model card says the pre-trained model is CC-BY-NC, a non-commercial license; only the code is Apache-2.0.
- Do I have to type what is said in my voice sample? No. If you leave out ref_text, the model transcribes the sample with Whisper by itself.
- Does it run on a Mac? Yes. The README gives an Apple Silicon install and says to use device_map="mps".
- Can I reuse a cloned voice later without the sample? Yes. Encode the sample once, save it as a file, and load it in later sessions.
- Are numbers read correctly? Write them as words, or pass normalize_text=True after installing the omnivoice[tn] extra.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
