Speech recognition systems have higher error rates for black people, study finds

No, you’re not just imagining it. Speech recognition systems developed by the likes of Amazon, Apple, Google, IBM, and Microsoft all have higher error rates when transcribing speech from black people than when doing so for white people. 

So finds a study published today in the Proceedings of the National Academy of Sciences, which notes that the increased prevalence of these systems in modern day life risks enshrining a new form of digital discrimination unless action is taken. 

“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,” wrote the study’s authors. “Our results point to hurdles faced by African Americans in using increasingly widespread tools driven by speech recognition technology.”

Researchers took 115 human-transcribed interviews — including conversations with 73 black speakers and 42 white speakers — and compared them to the versions produced by the tech giant’s speech-recognition tools. They discovered an “average word error rate” of almost double (0.35) when the system transcribed black speakers as compared to when it transcribed white speakers (0.19).” 

In other words, the systems worked noticeably worse for people of color. 

This is far from the first time we’ve seen evidence of bias embedded in the supposed technology of tomorrow. In December of last year, a federal study once again confirmed that facial-recognition tech is a biased mess. In October of 2017, we saw that Google’s text/sentiment-analysis tool exhibited signs of homophobia, racism, and antisemitism.  

However, it’s worth noting that with one quarter of U.S. adults claiming to have at least one smart speaker in their homes, the bias uncovered by today’s study is likely affecting tens of millions (if not more) people right now. 

SEE ALSO: Federal study confirms facial recognition is a biased mess

“With adoption of speech recognition systems likely to grow over time,” the study authors continue, “we hope technology firms and other participants in this field foreground the equitable development of these important tools.”

That’d sure be nice, wouldn’t it?