Vekol Piper edge (ckb)
Piper (VITS) 22.05 kHz ONNX edge model for Central Kurdish (Sorani). Runs on CPU / edge devices.
By Revge / Vekol ( ڕێڤگە / ڤەکۆڵ ).
Training data
Resumed from a Central Kurdish base, then trained on the Vekol voice plus teacher-distilled audio, with weak-letter balanced oversampling.
| Stage | Data | Hours | Clips |
|---|---|---|---|
| Base | DB1 + DB2 (Central Kurdish corpora) | 27.33 h | 11,694 |
| Fine-tune — real | Vekol voice | ~59 min | 397 sentences |
| Fine-tune — distilled | teacher (Vekol) | ~3 h | ~1,500 |
Weak letters (گ ڤ ح غ) are oversampled ×2; long-text tail-hallucination fixed in inference.
Usage
model.onnx + model.onnx.json with the Piper runtime, or the included vekol_tts.py helper.