SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR
SCRIBE is a diagnostic framework that breaks down Automatic Speech Recognition (ASR) errors into specific categories: lexical, punctuation, numeral, and domain-entity rates for Indic languages. It uses sandhi-tolerant alignment to accurately identify these errors. This method provides actionable signals for improving ASR systems beyond simple word error rates (WER), proving that Indic ASR is more reliable than standard metrics suggest.