Shenava — FastConformer-Hybrid Large (fa) — de-poisoned Phase B v4
115M EncDecHybridRNNTCTCBPE (RNNT+CTC), 16kHz. On-device offline Persian ASR for the VisualEars project.
Golden6669 (held-out gold, official Persian normalizer)
| head | WER | CER |
|---|---|---|
| RNNT | 7.29% | 1.63% |
| CTC | 7.92% | 1.87% |
vs prev best (B2) 8.02%/1.82%, vs cloud Gemini 6.49% — fully offline.
Recipe
De-poisoned 7,417h corpus (crap-classifier cut + gates + telephooney-CTC 534h), Phase A continued (→8.73%) + Phase B gold-anchor with 1,420 human corrections (→7.29%).