Racial disparities in automated speech recognition.

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.

Neighborhood effects on use of African-American Vernacular English.

African-American Vernacular English (AAVE) is systematic, rooted in history, and important as an identity marker and expressive resource for its speakers. In these respects, it resembles other vernacular or nonstandard varieties, like Cockney or Appalachian English. But like them, AAVE can trigger discrimination in the workplace, housing market, and schools. Understanding what shapes the relative use of AAVE vs. Standard American English (SAE) is important for policy and scientific reasons. This work presents, to our knowledge, the first experimental estimates of the effects of moving into lower-poverty neighborhoods on AAVE use. We use data on non-Hispanic African-American youth (n = 629) from a large-scale, randomized residential mobility experiment called Moving to Opportunity (MTO), which enrolled a sample of mostly minority families originally living in distressed public housing. Audio recordings of the youth were transcribed and coded for the use of five grammatical and five phonological AAVE features to construct a measure of the proportion of possible instances, or tokens, in which speakers use AAVE rather than SAE speech features. Random assignment to receive a housing voucher to move into a lower-poverty area (the intention-to-treat effect) led youth to live in neighborhoods (census tracts) with an 11 percentage point lower poverty rate on average over the next 10-15 y and reduced the share of AAVE tokens by ∼3 percentage points compared with the MTO control group youth. The MTO effect on AAVE use equals approximately half of the difference in AAVE frequency observed between youth whose parents have a high school diploma and those whose parents do not.