| --- |
| 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`). |
|
|
