Zion Mengesha 

Stanford University

Tuesday, March 29, 2022 4-5:30pm MST 

Automated speech recognition (ASR), which converts spoken language into text, is used in varied contexts such as medical translation services, crisis management, dictation, and closed captioning. While ASR systems have improved drastically due to advances in machine learning, research suggests that ASR does not work equally well for speakers of African American Vernacular English (Koenecke et al., 2021). Despite the relevance of sociolinguistic work to observed racial disparities in ASR, few studies have incorporated any insights from sociolinguistic theory in ASR research. For example, over the past several decades, sociolinguists have observed dialect discrimination, the auditory equivalent of visual racial discrimination, across institutions in classrooms, courtrooms, housing, and medicine (Williams, et al., 1971; Rickford, 1999; Purnell et al., 1999; Henderson, 2001; Nelson, 2002; Baugh, 2003; Grogger, 2011; Rickford & King, 2016; Jones et al., 2019). In today’s talk, I explore the behavioral and psychological consequences of being misheard by ASR through a 2-week long diary study of African Americans’ experiences with ASR failures. Through this lens, I provide a novel perspective on ASR errors by evaluating users’ experience of ASR errors as a form of dialect discrimination. The results demonstrate that ASR failures have a negative, detrimental impact on African Americans. Specifically, errors surface thoughts about identity, namely about race and geographic location—leaving African American users of voice technology feeling othered. As a result, ASR failures lead to anger, self-disappointment, and self-consciousness. These findings provide a window into the psychological consequences that AAVE speakers face when they experience dialect discrimination in day-to-day interactions with voice-AI. 

Join via Zoom: https://cuboulder.zoom.us/j/91010535502

(You need to join using your CU e-mail address)

(PASSCODE: GOOGLE)