phoonnx-mms

Meta MMS-TTS voices for phoonnx — VITS models covering ~1100+ languages. One voice per subfolder (model.onnx + native config.json + tokenizer). Mirror/conversion of facebook/mms-tts and community MMS finetunes. Part of the phoonnx community mirror.

License: cc-by-nc-4.0 (MMS-TTS). Phonemization is grapheme/uroman-based per the MMS pipeline. Loaded via the phoonnx voice index (engine: transformers).

Revisions

• 2026-09-18: the six hf_community__ylacombe__* voices were re-exported with scales as a graph input, so noise and rate reach the model at run time; the earlier export froze them. The six ghananlpcommunity__* configs now name the blank symbol (the tokenizer's pad token), so phoonnx feeds the same ids as the HuggingFace tokenizer. hf_community__NorHsangPha__mms-tts-shn-Shan keeps its earlier export: its source repository is no longer public, so it cannot be re-exported; its noise and rate are frozen. The 1,143 facebook__ voices already take the scales as inputs and are unchanged.