# Meta MMS-TTS — card for AI agents

Family of small VITS text-to-speech checkpoints, one per language, covering over 1,100 languages.

Category: Voiceover (text-to-speech) · maker Meta (FAIR) · checked 2026-10-25

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

## License
CC-BY-NC-4.0 (code and weights) — https://github.com/facebookresearch/fairseq/tree/main/examples/mms

- commercial: **no** (Non-commercial only)
  > The MMS code and model weights are released under the CC-BY-NC 4.0 license.
  (https://github.com/facebookresearch/fairseq/tree/main/examples/mms)
- attribution: **required** (Credit required)
  > The MMS code and model weights are released under the CC-BY-NC 4.0 license.
  (https://github.com/facebookresearch/fairseq/tree/main/examples/mms)
- in app: **no** (Cannot ship inside an app)
  > The MMS code and model weights are released under the CC-BY-NC 4.0 license.
  (https://github.com/facebookresearch/fairseq/tree/main/examples/mms)
- youtube: **no** (Not for monetized YouTube)
  > The MMS code and model weights are released under the CC-BY-NC 4.0 license.
  (https://github.com/facebookresearch/fairseq/tree/main/examples/mms)

## Rules to follow
- Non-commercial use only (CC-BY-NC 4.0).
- Credit Meta and link the license when you share the weights.

## How to connect
- Download only
- Repository: https://github.com/facebookresearch/fairseq
- Website: https://huggingface.co/facebook/mms-tts-eng

## Notes
One checkpoint per language (facebook/mms-tts-<iso639-3>); runs with VitsModel in Transformers 4.33+. Hugging Face API, Oct 25, 2026 (facebook/mms-tts-eng): 52721 downloads, 184 likes; GitHub facebookresearch/fairseq: 32212 stars.

## Numbers (AgentGid data)
- Gid Score: 46
- GitHub stars: 32,204

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

```
- Meta MMS-TTS (Voiceover (text-to-speech)). Family of small VITS text-to-speech checkpoints, one per language, covering over 1,100 languages. Price: free; non-commercial only; credit required; cannot ship inside an app.
  Connect: download from https://huggingface.co/facebook/mms-tts-eng (no API).
  Rule: Non-commercial use only (CC-BY-NC 4.0).
  Rule: Credit Meta and link the license when you share the weights.
  License: https://github.com/facebookresearch/fairseq/tree/main/examples/mms (checked 2026-10-25; details: https://agentgid.com/assets/mms-tts/)
```

Not verified: Download and like counts are for the English checkpoint only, not the whole family. The fairseq repo is archived (read-only) on GitHub.

## Links
- AgentGid page: https://agentgid.com/assets/mms-tts/
- Is it free for commercial use? https://agentgid.com/assets/mms-tts/commercial-use/
- Source: https://huggingface.co/facebook/mms-tts-eng
- Source: https://github.com/facebookresearch/fairseq/tree/main/examples/mms
- Source: https://huggingface.co/api/models/facebook/mms-tts-eng


## Meta MMS-TTS: free speech voices for 1,100+ languages

MMS-TTS is a family of free text-to-speech voices from Meta, one small voice file per language, covering over a thousand languages. You or your AI assistant run it on your own computer to turn written text into spoken audio. It is for non-commercial projects only, such as research, teaching or hobby work.

### Where to use it
- Speech in languages paid voices skip: Hear text read aloud in a language that big voice services do not offer, for a study, a class or a community project. Each language has its own voice, picked by its three-letter code.
- Audio files from a script: Turn a list of sentences into sound files on your own computer, for example short prompts for a language lesson or a museum audio guide that is not sold.
- Text in non-Latin scripts: Languages written in other alphabets, such as Korean or Arabic, need one extra free package that turns the text into Latin letters first. The voice then reads it as usual.
- Faster or slower reading: Make the voice read faster or slower, for example slower for language learners. One setting changes the speed of the whole clip.
- Read-aloud on a hobby site: Add a listen button to a free, non-commercial site, such as a community dictionary or a school project. Make the audio files ahead of time and serve them like any other file. (idea, not from the official pages)
- Try a language before buying a voice: Hear how a new language sounds in your app's flow before you choose a paid voice for the real product. Swap in the paid voice before launch. (idea, not from the official pages)

### How to start
- Install the Transformers library (version 4.33 or newer) with PyTorch, in a fresh project environment.

```
pip install --upgrade transformers accelerate
```

- Load the voice for your language by its code (facebook/mms-tts-<code>) and turn text into a waveform.

```
model = VitsModel.from_pretrained("facebook/mms-tts-eng")
tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-eng")
```

- Save the result as a WAV file at the model's sampling rate.

```
scipy.io.wavfile.write("techno.wav", rate=model.config.sampling_rate, data=output.float().numpy())
```


### License
Free only for non-commercial use, under CC BY-NC 4.0. Credit Meta, link the license and say if you changed anything. Paid apps, client work and monetized videos are not allowed.

### Good to know
- No commercial use: the code and the voice files are under a non-commercial license.
- The same text can sound a little different each run. Set a fixed seed when you need the exact same audio again.
- There is one voice per language, so you cannot pick between several voices or clone your own from these files alone.
- The original Meta code project is archived and read-only, so expect no new fixes there; the Transformers version is the easier path.
- Languages in non-Latin scripts need the extra uroman package before the text goes in.

Checked 2026-11-10 against: https://huggingface.co/facebook/mms-tts-eng, https://huggingface.co/facebook/mms-tts-eng/raw/main/README.md, https://huggingface.co/api/models/facebook/mms-tts-eng/tree/main, https://huggingface.co/facebook/mms-tts, https://github.com/facebookresearch/fairseq/blob/main/examples/mms/README.md, https://github.com/facebookresearch/fairseq, https://huggingface.co/docs/transformers/model_doc/vits, https://creativecommons.org/licenses/by-nc/4.0/, https://pypi.org/pypi/transformers/json

### Mistakes to avoid
- Don't: Using MMS-TTS voices in a paid app, an ad or a client project → Do: Use it only for research, teaching or hobby work; pick a commercially licensed voice (for example Chatterbox, MIT) for products (The weights are under a non-commercial license.)
- Don't: write("out.wav", rate, waveform.numpy()) with the (1, samples) tensor as is → Do: write("hello.wav", sampling_rate, audio_data.squeeze()) (The docs squeeze the batch dimension before saving, so the file holds one channel of audio.)
- Don't: Generating twice and expecting identical audio → Do: Call set_seed(555) (any fixed number) before generating (VITS adds randomness to rhythm, so output changes between runs.)
- Don't: Feeding Korean, Arabic or other non-Latin text without uroman installed → Do: pip install -U uroman, then check tokenizer.is_uroman (Those voices were trained on romanized text.)
- Don't: pip install --upgrade transformers with no version, as the model card shows → Do: Pin a tested Transformers version (4.33 or newer) in requirements (The model needs 4.33+, and an unpinned upgrade can change behaviour between runs of your project.)

### Questions
- Can I use MMS-TTS voices in a paid app or a monetized video? No. The code and the voice files are under CC BY-NC 4.0, which does not allow commercial use.
- How do I get a voice for another language? Each language has its own checkpoint named by its ISO 639-3 code; Meta publishes the list of codes for the 1107 languages.
- Why does the same sentence sound slightly different each time? The model picks speech rhythm with some randomness; set a fixed seed to get the same audio again.
- Which software version do I need? Hugging Face Transformers 4.33 or newer runs MMS-TTS.

---
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.
