The technology behind automated speech recognition systems at some of the nation’s top tech firms are twice as likely to misinterpret words spoken by African-Americans compared to those spoken by whites, according to a new study out of Stanford University.
The report, published on March 23, identified significant racial disparities in systems from Amazon, Apple, Microsoft, Google and IBM, which each made far more errors with users who are Black.
On average, the systems misunderstood white people’s words 19 percent of the time. That figure rose to 35 percent when it came to African-Americans’ speech. Researchers also ran additional tests and found that over 20 percent of audio transcriptions of Black voices were unreadable, compared to just 2 percent for white voices.
“Error rates were highest for African American men,” according to Stanford News, which cited the study, “and the disparity was higher among speakers who made heavier use of African American Vernacular English.”
The culprit? Researchers believe the technology may be flawed because the systems on which they’re trained don’t use diverse data and rely too heavily on English spoken by white people. Allison Koenecke, a doctoral candidate and lead author of the report, said the findings point to a need for inclusion.
“One should expect that U.S.-based companies would build products that serve all Americans,” she said. “Right now, it seems that they’re not doing that for a whole segment of the population.”
For the study, Koenecke and her team of researchers tested more than 2,000 speech samples across each of the five systems, including voice interviews from whites and African-Americans; each system had error rates that were twice as high for Black users as they were for white users,
Stanford News notes that in today’s tech-driven world, companies rely on voice automation for a number of things, from screening prospective hires to courtroom transcriptions. Researchers warn the disparities could have far-reaching impacts if unaddressed, especially since tech giants like Apple are not legally obligated to disclose what powers their products or how they’re supposed to work.
“A more equitable approach would be to include databases that reflect a greater diversity of the accents and dialects of other English speakers,” the report notes.
The study also suggests auditing new technologies, like speech recognition, for potential bias that could further exclude already marginalized groups.