SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR

TL;DR

SCRIBE is a diagnostic framework that decomposes ASR errors into categories, addressing limitations of WER, and provides open models and benchmarks for Indic languages.

Contribution

It introduces SCRIBE, a novel error decomposition method for Indic ASR, along with open models and benchmarks for Hindi, Malayalam, and Kannada.

Findings

SCRIBE aligns with expert judgment better than WER

Provides categorical error rates for lexical, punctuation, numeral, and domain-entity errors

Open-weight transcription models released for multiple Indic languages

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. Word error rate (WER) fails on two fronts: it collapses distinct error categories into a single scalar, and it structurally penalizes agglutinative languages where valid sandhi merges inflate scores. We introduce SCRIBE, a diagnostic framework that provides categorical error decomposition into lexical, punctuation, numeral, and domain-entity rates through sandhi-tolerant alignment with domain vocabulary injection. Human validation confirms SCRIBE aligns with expert judgment where WER does not. We release SCRIBE, an LLM curation pipeline, benchmarks, and open-weight rich transcription models for Hindi, Malayalam, and Kannada.