afri-vits-bases

Starting checkpoints for fine-tuning VITS TTS in African languages, one folder per language code (361 folders). Each folder is a Meta MMS-TTS generator whose text vocabulary has been expanded with the IPA symbols produced by africa-g2p (about 190 symbols covering its 750 languages), so IPA text for any African language can be fed in.

Use them through AfriSpeech/afri-vits-finetuner, which maps any africa-g2p language to its best base (afri-vits pick-base <lang>) and runs LoRA or full fine-tuning on your data.

Important

  • These models do not speak new languages on their own. The added symbol embeddings start from the nearest existing sound or the mean embedding plus noise and are meant to be learned during fine-tuning. Without fine-tuning a base behaves like the original MMS model for its own language.
  • Generator only. The discriminator needed for GAN training is fetched from facebook/mms-tts by the fine-tuner.
  • Each folder's ipa_expansion.json lists the original vocab size, the symbols added and how each row was initialised.
  • krs is absent: its upstream MMS repo has no weights.

Licence

Derivative of Meta's MMS-TTS checkpoints: CC BY-NC 4.0, non-commercial use only, with attribution to the MMS project.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for AfriSpeech/afri-vits-bases

Base model

facebook/mms-tts
Finetuned
(15)
this model