When evaluating speech recognition for Indic languages, use SN-WER alongside standard WER to avoid inflated error rates caused by script differences rather than actual recognition failures.
This paper introduces Script-Normalized WER (SN-WER), a new evaluation metric for speech recognition in languages that use multiple writing systems. The metric transliterates both the reference and predicted text into a standard script before calculating error rates, preventing the metric from penalizing correct recognitions that are just written in different scripts (like romanized vs.