Confidence-Aware Training Boosts Medical ASR for Dravidian Languages

A new confidence-aware training framework improves automatic speech recognition (ASR) for low-resource Dravidian languages Telugu and Kannada in medical domains. The approach integrates real and synthetic speech via a hybrid confidence mechanism combining static perceptual/acoustic similarity metrics with dynamic model entropy. Two aggregation strategies—fixed-weight and learnable-weight—guide sample weighting during training. Evaluation on medical datasets with real recordings and TTS-generated speech, plus a 5-gram KenLM language model for post-decoding correction, shows performance gains.