whisper-small fine-tuned on Bengali (AIMScribe RA assessment)
Full fine-tune of openai/whisper-small on an 80-minute speaker-disjoint subset of OpenSLR 53 (Bengali read speech).
Settings: lr 1e-5, batch 8 x 2, 800 steps, fp16, seed 42; best checkpoint by dev WER (step 600, 6 held-out speakers).
| OpenSLR 53 test (15 unseen speakers) | WER | CER |
|---|---|---|
| whisper-small zero-shot | 144.1 | 135.7 |
| this model | 66.4 | 21.4 |
A small-data research baseline, not a production model. Code, evaluation and the full report are in the assessment submission (Task 2).