F5-TTS Yoruba

Fine-tuned F5-TTS (F5TTS_v1_Base) for Yoruba zero-shot TTS.

Files

model_last.pt — EMA-pruned inference checkpoint
vocab.txt — custom vocab (base + Yoruba graphemes / combining tone marks)

Inference

Text MUST be normalized the same way it was at training: NFC, junk-char cleanup, and lowercased. Reuse text_clean() from the training prep.py.

from f5_tts.api import F5TTS
m = F5TTS(model="F5TTS_v1_Base", ckpt_file="model_last.pt", vocab_file="vocab.txt", use_ema=True)
wav, sr, _ = m.infer(ref_file="ref.wav", ref_text=text_clean("..."),
                     gen_text=text_clean("..."), nfe_step=32)

If outputs are noise, retry with use_ema=False.

Known limitation

Cultural named entities (e.g. Òrìṣà names: Ṣàngó, Ọya, Ọba) are under-articulated, a data-sparsity / representation effect. Addressing it needs targeted name data or a phoneme (G2P) frontend rather than more epochs.

Downloads last month: 104