Bias in Automatic Speech Recognition: The Case of African American Language

Abstract

Research on bias in artificial intelligence has grown exponentially in recent years, especially around racial bias. Many modern technologies which impact people’s lives have been shown to have significant racial biases, including automatic speech recognition (ASR) systems. Emerging studies have found that widely-used ASR systems function much more poorly on the speech of Black people. Yet, this work is limited because it lacks a deeper consideration of the sociolinguistic literature on African American Language (AAL). In this paper, then, we seek to integrate AAL research into these endeavors to analyze ways in which ASRs might be biased against the linguistic features of AAL and how the use of biased ASRs could prove harmful to speakers of AAL. Specifically, we (1) provide an overview of the ways in which AAL has been discriminated against in the workforce and healthcare in the past, and (2) explore how introducing biased ASRs in these areas could perpetuate or even deepen linguistic discrimination. We conclude with a number of questions for reflection and future work, offering this document as a resource for cross-disciplinary collaboration.

Accuracy pecking order – How 30 AI detectors stack up in detecting generative artificial intelligence content in university English L1 and English L2 student essays

“…Adamson, 2023;Gillham, 2024). This finding resonates, in a different but related scenario, with the view that some studies have established that currently available automatic speech recognition technologies poorly detect, if any, and discriminate against the English spoken by Black people, especially African American Language (AAL), thereby exposing their racial bias and demographic discrimination against this type of English (Martin & Wright, 2023). Linguistic and racial biases are but two of the instances of bias that GenAI models, and not just AI detection models, have to contend with in their everyday deployment.…”

Essentialist characterizations of language are an obstacle to accuracy, progress, and justice in science

“…Though the actual abilities of these tools may be overhyped 92,93 , we may be witnessing the emergence of a new benchmark against which human language use is measured. Concerningly, the biases embedded in these tools, which take the linguistic practices of socially powerful groups as the baseline for NATIVE, "proficient," or "human-like" language use, reinforce linguistic discrimination and contribute to the dehumanization of marginalized communities 92,94,95 .…”

Mediated Bricolage and the Sociolinguistic Co-Construction of No Sabo Kids

“…For example, Carter and Callesano (2018) note that different national dialects of Spanish (Cuban, Colombian, and Peninsular) are associated with different levels of occupational prestige and socioeconomic status. African American English (AAE) has long been studied in terms of materialized perceptions, for example in the workplace (McCluney et al 2021), the court system (Rickford and King 2016), and automatic speech recognition programs (Martin and Wright 2023). As such, the perception of minoritized linguistic varieties-and the bodies that produce them-materialize as not eligible for higher paying employment, not credible in legal environments, and disadvantaged within technological systems, underscoring the ways in which marginalized Latinx bilingual language practices are policed.…”