Racial disparities in automated speech recognition

Source: PNAS 117(14):7684-7689 (Koenecke, Nam, Lake, Nudell, Quartey, Mengesha, Toups, Rickford, Jurafsky, Goel; Stanford) · 2020-03-23

Automated speech recognition (ASR) systems are now used in a variety of applications to convert spoken language to text, from virtual assistants, to closed captioning, to hands-free computing. By analyzing a large corpus of sociolinguistic interviews with white and African American speakers, we demonstrate large racial disparities in the performance of five popular commercial ASR systems. Our results point to hurdles faced by African Americans in using increasingly widespread tools driven by speech recognition technology. More generally, our work illustrates the need to audit emerging machine-learning systems to ensure they are broadly inclusive.

Automated speech recognition (ASR) systems, which use sophisticated machine-learning algorithms to convert spoken language to text, have become increasingly widespread, powering popular virtual assistants, facilitating automated closed captioning, and enabling digital dictation platforms for health care. Over the last several years, the quality of these systems has dramatically improved, due both to advances in deep learning and to the collection of large-scale datasets used to train the systems. There is concern, however, that these tools do not work equally well for all subgroups of the population. Here, we examine the ability of five state-of-the-art ASR systems—developed by Amazon, Apple, Google, IBM, and Microsoft—to transcribe structured interviews conducted with 42 white speakers and 73 black speakers. In total, this corpus spans five US cities and consists of 19.8 h of audio matched on the age and gender of the speaker. We found that all five ASR systems exhibited substantial racial disparities, with an average word error rate (WER) of 0.35 for black speakers compared with 0.19 for white speakers. We trace these disparities to the underlying acoustic models used by the ASR systems as the race gap was equally large on a subset of identical phrases spoken by black and white individuals in our corpus. We conclude by proposing strategies—such as using more diverse training datasets that include African American Vernacular English—to reduce these performance differences and ensure speech recognition technology is inclusive.

There is worry, however, that speech recognition systems suffer from racial bias (1–4), a problem that has recently come to light in several other advancing applications of machine learning, such as face recognition (5, 6), natural language processing (7–11), online advertising (12, 13), and risk prediction in criminal justice (14–17), healthcare (18, 19), and child services (20, 21). Here, we assess racial disparities in five commercial speech-to-text tools—developed by Amazon, Apple, Google, IBM, and Microsoft—that power some of the most popular applications of voice recognition technology.

Our analysis is based on two recently collected corpora of conversational speech. The first is the Corpus of Regional African American Language (CORAAL) (22), a collection of sociolinguistic interviews with dozens of black individuals who speak African American Vernacular English (AAVE) (23–25) to varying degrees. These interviews were conducted at three US sites: Princeville, a rural, nearly exclusively African American community in eastern North Carolina; Rochester, a moderate-sized city in Western New York; and the District of Columbia. The second dataset we use is Voices of California (VOC) (26), an ongoing compilation of interviews recorded across the state in both rural and urban areas. We focus our analysis on two California sites: Sacramento, the state capitol; and Humboldt County, a predominately white rural community in Northern California.

In both datasets, the interviews were transcribed by human experts, which we use as the ground truth when evaluating the performance of machine transcriptions. The original recorded interviews contain audio from both the interviewer and the interviewee. Our study is based on a subset of audio snippets that exclusively contain the interviewee and are 5 to 50 s long. We match these snippets across the two datasets based on the age and gender of the speaker and the duration of the snippet. After matching, we are left with 2,141 snippets from each dataset, with an average length of 17 s per snippet, amounting to 19.8 total hours of audio. In the matched dataset, 44% of snippets were of male speakers, and the average age of speakers was 45 y.

We assess the performance of the ASR systems in terms of the word error rate (WER) (27), a standard measure of discrepancy between machine and human transcriptions. Formally, WER is defined as:

where S, D, and I denote the number of word substitutions, deletions, and insertions between the machine and ground-truth transcriptions, respectively, and N is the total number of words in the ground truth. A higher WER thus indicates a greater difference between the two transcriptions and hence worse ASR performance in our setting.

We start by computing the average word error rates for machine transcriptions across our matched audio snippets of white and black speakers. For each of the five commercial ASR systems we examine, Fig. 1 shows that the average WER for black speakers is substantially larger than the average WER for white speakers. For example, for Microsoft’s ASR, which has the best overall performance, the WER for black speakers is 0.27 (SE: 0.004) compared with 0.15 (SE: 0.003) for white speakers. Furthermore, for Apple, whose ASR has the worst overall performance, the WERs for black and white speakers are 0.45 (SE: 0.005) and 0.23 (SE: 0.003), respectively.* Despite variation in transcription quality across systems, the error rates for black speakers are nearly twice as large in every case. Averaging error rates across ASR services yields an aggregate WER of 0.35 (SE: 0.004) for black speakers versus 0.19 (SE: 0.003) for white speakers.