Arnav Jagetiya
.
Download PDF
http://doi.org/10.37648/ijrst.v16i03.006
Automatic speech recognition (ASR) systems are evaluated by an aggregate word error rate (WER) that conceals which speakers a model fails on. In this paper, we measure that hidden disparity for OpenAI’s Whisper, comparing its transcription accuracy on Indian-accented and American-accented English. Sixty speakers from each group read an identical passage from the Speech Accent Archive, so accent is the only variable between conditions. We transcribed each recording with four Whisper sizes (tiny, base, small and medium), computed per-clip WER using jiwer, and tested group differences with a one-sided Mann-Whitney 𝑈 test. We also ran the same corpus through Meta’s MMS-1B-all and NVIDIA’s Parakeet to test whether the gap is specific to Whisper.
Disclaimer: Indexing of published papers is subject to the evaluation and acceptance criteria of the respective indexing agencies. While we strive to maintain high academic and editorial standards, International Journal of Research in Science and Technology does not guarantee the indexing of any published paper. Acceptance and inclusion in indexing databases are determined by the quality, originality, and relevance of the paper, and are at the sole discretion of the indexing bodies.
