Automatic Speech Recognition Systems, Spoken Corpora, and African American Language: An Examination of Linguistic Bias and Morphosyntactic Features

Abstract

In recent years, the use of automatic speech recognition (ASR) has expanded far beyond the bounds of computer science research to arenas that impact everyday life. Yet, while expanding implementation of ASRs may be signaling a new era of human-computer interaction, racial bias in ASR systems threatens to negate the beneficial outcomes they could produce for speakers of color within the United States, in particular Black speakers. However, while this is the case, beyond a select few studies, little has been done to survey the extensive potential concrete harms biased ASR systems could present for African American users, to understand the impact of the presence of specific African American Language (AAL) linguistic features within speech data on ASR performance, or to investigate the nature of AAL within speech datasets used for ASR development and evaluation. This research aims to bridge this gap by exploring linguistic bias in ASR systems centered on bias against African Americans, particularly those users whose linguistic repertoires contain the linguistic features of AAL. In particular, the studies contained here explore the potential impacts that ASR systems biased against the speech of various African American communities and individuals could have, evaluate the performance of widely used ASR services on the recorded speech of African Americans containing canonical morphosyntactic features of AAL, and investigate the presence of AAL in standard spoken corpora used most widely in ASR research and in the evaluation of commercial ASR systems through a consideration of AAL morphosyntactic features. The results find that biased ASRs have the potential to cause serious harm to the lives of speakers of AAL, that current major ASR systems perform much more poorly when presented with speech data containing AAL morphosyntactic features, and that the frequency, dispersion, and contexts of AAL morphosyntactic features within spoken corpora widely used for ASR development and evaluation do not reflect the nature of those features as found within a spoken reference corpus of AAL. Taken together, this research points to a pressing need to address bias in ASR systems to ensure equitable treatment for African American users, particularly speakers of AAL.