README.md

---
license: apache-2.0
tags:
- tts
- arabic
- kokoro
- styletts2
- checkpoint
---

# Sofelia TTS — Training Checkpoints (private)

Raw StyleTTS2 Stage-2 checkpoints for the Palestinian Arabic Kokoro fine-tune.
These are training-state checkpoints (`{'net': {...module.prefixed...}, 'optimizer', 'epoch', ...}`),
**not** the inference model. For inference use the public model
[hamdallah/Sofelia-TTS-82M](https://huggingface.co/hamdallah/Sofelia-TTS-82M).

## Checkpoints

| File | What | Val |
|---|---|---|
| `checkpoints/sirin_eliaa_final_epoch2nd.pth` | **Eliaa** single-voice final (powers the public model) | 0.463 (Sirin val set) |
| `checkpoints/multispeaker_final_epoch2nd.pth` | multispeaker base (Kore + Sirin) | 0.446 (mixed val set) |
| `checkpoints/stage1_first_stage.pth` | Stage-1 acoustic checkpoint | — |

## Provenance

Kokoro-82M → Stage 1 (multispeaker, Palestinian Arabic) → Stage 2 (multispeaker)
→ Stage 2 continuation on the single human voice (Eliaa/Sirin), stopped at the
validation plateau / overfit point. `training/` holds the configs, the dialect
frontend (`prepare_arabic.py`, `ar_lexicon.json`), and the voicepack/inference
scripts needed to reproduce or extend.

To convert a checkpoint to the public Kokoro inference format, strip the
`module.` prefixes from each `net` component and load via `kokoro.KModel`
(see `training/test_inference_arabic.py`).