SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR
The article introduces SCRIBE, a diagnostic framework designed for evaluating automatic speech recognition (ASR) in Indic languages. SCRIBE addresses limitations of traditional word error rate (WER) metrics by providing a detailed error decomposition. The framework has been validated by human experts and includes open-weight transcription models for Hindi, Malayalam, and Kannada.
- ▪SCRIBE offers a categorical error decomposition into lexical, punctuation, numeral, and domain-entity rates.
- ▪Traditional WER metrics fail to account for distinct error categories and agglutinative language structures.
- ▪Human validation confirms that SCRIBE aligns better with expert judgment compared to WER.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computation and Language arXiv:2605.20712 (cs) [Submitted on 20 May 2026] Title:SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR Authors:Kavya Manohar, Arghya Bhattacharya, Kush Juvekar, Kumarmanas Nethil View a PDF of the paper titled SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR, by Kavya Manohar and 3 other authors View PDF HTML (experimental) Abstract:Automatic speech recognition replaces typing only when correction costs less than manual entry, a threshold determined by error types, not counts: fixing a misrecognized domain term costs far more than inserting a comma.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
