africa-vc

Voice conversion for African languages, fine-tuned from Seed-VC on AfriSpeech/multivoice-synthetic-speech — 17,010 clips, 567 African languages, 30 voices, 38.7 hours.

What it does

It converts who is speaking, not what is said. The words and the language come from the source audio; only the voice is replaced. It cannot make a model speak a language — to get Twi out, the source must already be Twi.

Because it was trained on all 30 voices rather than one, the target voice is a runtime choice: the reference clip selects it, and every one of the 30 is reachable from this single checkpoint.

Use

pip install git+https://github.com/AfriSpeech/africa-vc
africa-vc voices
africa-vc convert --source speech.wav --voice Sulafat --checkpoint ft_model.pth

Training

Base	Seed-VC 51383efd9210
Config	config_dit_mel_seed_uvit_whisper_small_wavenet.yml
Data	17,010 clips, 566 languages, 30 voices
Steps	20,000
Batch size	2
Hardware	Modal A100-40GB

Caveats

The training audio is synthetic, generated with Google Gemini. The model learns those voices, and where the synthesiser mispronounced a language it learns that too. With 30 speakers from one TTS family there is a real risk of overfitting to synthetic timbre and generalising less well to human reference clips — check against real speech before relying on it.

Quality varies enormously by language. The voices were built for widely spoken languages and were asked to read hundreds of others.