F5-TTS Hebrew v2

Hebrew text-to-speech model fine-tuned from F5-TTS on 158 hours of Hebrew speech.

Model Details

• Base: F5-TTS v1 Base (non-autoregressive)
• Steps: 320,000
• Data: 68,569 segments (~158h) from SASPEECH Gold/Auto, FLEURS, Hebrew Speech Campus
• All data re-vocalized with Phonikud G2P (stress marks included)
• Hebrew-filtered: non-Hebrew segments removed
• Output: 24kHz mono WAV
• 58 built-in voices with voice cloning support

Critical Usage Notes

1. Use model_state_dict, NOT ema_model_state_dict when loading from .pt checkpoints
2. Override text_num_embeds to match vocab_size (default 256 is wrong)
3. Use the included vocab.txt — char-to-index mapping must match exactly
4. Call model.sample() directly — F5TTS API/CLI are broken for fine-tuned models

Full Project

github.com/yzamari/hebAudio — complete Hebrew TTS system with Web UI, 58 voices, preprocessing pipeline, and training guide.