gijl/level-tts-ar

Arabic-only (multi-dialect) zero-shot TTS / voice-cloning model, trained from scratch with level.py (flow-matching DiT, F5-TTS family). Preset: base. Last checkpoint: step 5649.

latest.pt - full training state (model, optimizer, EMA, step) used to resume training.
ema_model.pt - EMA weights in fp16 for inference.
vocab_ar.json - character vocabulary + dialect tags.

Dialect codes: msa, eg, sa, bh, kw, ae, qa, om, ye, iq, sy, lb, ps, jo, sd, ly, tn, dz, ma, mr, unk. Mel: 24000 Hz, 100 bands, hop 256 (pair with a matching Vocos vocoder).

Downloads last month: -; Downloads are not tracked for this model. How to track