VAYAS: An Age-Stratified Zero-Shot Audit of Hindi Speech Recognition

We present a zero-shot audit of four automatic speech recognition (ASR) systems for Hindi, examining whether word- and character-error rates differ by speaker age under a matched-cohort design. Common Voice Hindi, the most widely used open Hindi ASR benchmark, turns out to have almost no confirmed elderly speakers once its self-reported age field is actually counted (2 in their 60s, 0 older, out of 369 speakers). We instead use IndicVoices-R Hindi, a public spontaneous-speech corpus with genuine elderly (age 60+) representation, comparing 50 elderly speakers against a duration-matched, same-corpus control cohort across four systems: OpenAI Whisper large-v3, Meta MMS-1B, and two India-specialized models, AI4Bharat's IndicConformer and IndicWhisper. IndicConformer achieves the lowest error rates by a wide margin, though we flag a serious, unresolved training-data contamination risk on this result. IndicWhisper is the weakest performer, partly explained by a decoder repetition-loop failure mode. Age-cohort effect sizes are small across all four systems (Cohen's d < 0.2 in every case); one nominally significant elderly disadvantage (IndicWhisper) does not survive correction for multiple comparisons and is below this sample's minimum detectable effect size, so we report it as suggestive, not confirmed. We release the full zero-shot audit pipeline, matched-cohort sampling code, and per-clip results.