Symmetris-labs/tts-mms

MMS-TTS for sherpa-onnx (1000+ languages)

Meta's MMS-TTS checkpoints (part of the Massively Multilingual Speech project), re-exported to the ONNX format used by k2-fsa/sherpa-onnx's OfflineTts (VITS family, character-level tokenization) - offline, on-device text-to-speech for over 1000 languages.

Usage with sherpa-onnx

Each language folder (ISO 639-3 code, e.g. tam/, heb/, eng/) contains:

• model.onnx - re-exported from the original .pth checkpoint via vits-mms.py, the official conversion script from the k2-fsa/sherpa project.
• tokens.txt - character vocabulary, generated from the original vocab.txt.

Load it like any other sherpa-onnx VITS/MMS model - e.g. in the sherpa_onnx Dart/Flutter package:

final config = OfflineTtsConfig(
  model: OfflineTtsModelConfig(
    vits: OfflineTtsVitsModelConfig(
      model: '$languageDir/model.onnx',
      tokens: '$languageDir/tokens.txt',
    ),
  ),
);
final tts = OfflineTts(config);

Same idea applies to the Python, C++, Kotlin, Swift, C#, Go, and Rust sherpa-onnx APIs - point them at model.onnx/tokens.txt the same way you would any other VITS voice.

Related repos

• Symmetris-labs/tts-piper - Piper voices re-hosted for CORS-friendly access, same sherpa-onnx OfflineTts format.
• Symmetris-labs/tts-kokoro - Kokoro multi-speaker voices, same format.

License

CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0) - same license as every original facebook/mms-tts-* checkpoint. Non-commercial use only. Attribution: Meta AI, Massively Multilingual Speech project.

Used by bible_with_me, a free app with no monetization.