Start Date: 12/3/2026 3:00 PM EST
End Date: 12/3/2026 4:00 PM EST
Organization Name:
Linguistic Society of America (LSA)
Contact:
Meet the Authors Webinar
Dialect Bias in Automatic Speech Recognition: Analysis of Appalachian English
Thursday, December 3, 2026, 3:00-4:00 PM ET
Speakers: Li-Fang Lai, Joe Salmons, Jennifer Cramer, Derek Denis, Charles Carson
Voice AI is now part of everyday life, yet automatic speech recognition (ASR) systems perform unevenly across dialects. In this encore presentation of our article (Lai, van Hell, and Lipski 2025), I examine how one ASR system processed Appalachian English, a regional variety that remains understudied in ASR research and is frequently the target of negative stereotypes. Comparing read speech from Eastern Tennessee (Southern Appalachian) and Central Pennsylvania (non-Southern Appalachian) speakers, we found consistently lower vowel recognition accuracy for Southern Appalachian speech, with half of the Southern errors traceable to the Southern Vowel Shift (SVS), including the distinctly Appalachian monophthongization of /a?/ before voiceless consonants. A spot-check with a newer end-to-end system suggests that these errors persist. I then extend the published work with new exploratory analysis of the speakers' own interviews, asking whether their attitudes toward their hometown and their speech relate to the SVS-related errors the system makes. Speakers who described adjusting or suppressing their speech still produced errors largely tied to vowel shifts. This pattern may indicate that while speakers can consciously change the words and features they notice, vowel shifts remain largely below conscious control, and these are precisely what the machine mishears. Throughout, I highlight what linguists can contribute to building speech technologies that serve linguistically diverse speakers.
