IndicConformer-120M Bengali, clinical-vocabulary fine-tune (AIMScribe RA assessment)
Fine-tune of AI4Bharat's indicconformer_stt_bn_hybrid_ctc_rnnt_large (needs AI4Bharat's NeMo fork, branch nemo-v2).
Use CTC decoding: only the CTC branch was trained.
Data: 1,012 synthetic clinical sentences (drug + dose + frequency, tests, diagnoses, replies) voiced by IndicF5 cloning 64 OpenSLR 53 speakers, after a QC step that dropped clips with CER > 50 %, mixed 1:1 with real OpenSLR 53 speech. No test sentence and no word from the dev consultations is in the training text. Training: 1,500 steps, batch 8, lr 5e-5, best checkpoint by dev WER (6 held-out OpenSLR speakers: 17.8 → 15.5).
| (CTC decoding) | zero-shot 120M | this model |
|---|---|---|
| Real role-play consultations, CER | 11.1 | 9.4 |
| Synthetic clinical test (600 clips), CER | 4.6 | 3.6 |
| Clinical term accuracy, all terms | 51.6 | 70.7 |
| ... terms seen in training / unseen | 50.9 / 44.1 | 69.7 / 61.7 |
A research model from a 7-day assessment. It did not beat IndicConformer 600M (consultation CER 7.2), so that model is used in the final pipeline. Not for clinical use.