timothydillan/f5-tts-balinese

F5-TTS — Balinese (v0)

A full fine-tune of F5-TTS (F5TTS_v1_Base) on ~29 min of single-speaker Balinese speech, part of Open Indonesia Models. Character-input (no G2P), zero-shot voice cloning. v0 baseline.

• model_last.pt — fine-tuned checkpoint (load via F5-TTS --ckpt_file).
• vocab.txt — character vocab (base 2545 vocab; Balinese romanization adds no new chars).

Usage

pip install f5-tts
f5-tts_infer-cli --model F5TTS_v1_Base --ckpt_file model_last.pt --vocab_file vocab.txt \
  --ref_audio <clean_same_speaker.wav> --ref_text "<ref transcript>" \
  --gen_text "Rahajeng semeng" --output_dir out --output_file demo.wav

Known v0 limitation

Words ending in "e" can drop the final vowel (Balinese final-e is often a reduced schwa /ə/; character input doesn't disambiguate). A Balinese G2P / data fix is planned.

License & attribution

Base F5-TTS weights: CC BY-NC 4.0 (non-commercial). Trained on a Balinese TTS dataset (single speaker). For research / non-commercial use.