When it comes to technology, we know that there is racial bias. And unfortunately, automatic speech recognition (ASR) is no different. But Howard University and Google are trying to change that.
They are working to build “an African American English speech dataset that will then be available to others looking to improve speech technology,” reports NPR.
Gloria Washington, an associate professor of computer science at Howard, is the principal investigator for the Elevate Black Voices project. Washington told NPR about the data collection process, and how they understand there are regional dialects. Thus, they’re “going to the Deep South, Alabama and also [the] Houston area so we can collect some audio segments so that we can use them later on down the line.”
Washington’s long term goal: “a version of a Siri or Alexa that has the ability to code switch, where you can allow it to speak naturally African American English to you, and you can be comfortable interacting with it.”
The harmful impacts of dialect discrimination became a public topic of discourse after the 1979 landmark federal court case Martin Luther King Junior Elementary School Children v. Ann Arbor School District. “Eleven African American children were wrongly placed in special education when Language Arts teachers and standardized tests failed to capture their command of English.” The Michigan Supreme Court ruled that the failure of the school to recognize the Black students’ speech patterns was a violation of federal law.
Clearly, misunderstandings in Black dialects is still a problem forty-four years later. In 2020, a study published in the journal, Proceedings of the National Academy of Sciences, proved that there were racial disparities in ASR. Their study found that typically, the systems misunderstood 35 percent of Black words versus 19 percent of whites.
Stanford University researchers conducted the study, and also found that approximately “2 percent of audio snippets from white people were considered unreadable by these systems…That rose to 20 percent with Black people.” The researchers conjectured this was because of a flaw in the machine learning systems.
This is yet another sign of bias as artificial intelligence (A.I.) become more and more prevalent parts of our day to day lives. Other research has demonstrated the problems with facial recognition software as it relates to law enforcement. “[T]hey can be far less accurate when trying to identify women and people of color.”
A statistics professor at New York University who studies discrimination in new technologies is frustrated about this bias. Ravi Shroff said, “I don’t understand why there is not more due diligence from these companies before these technologies are released,” continuing, “I don’t understand why we keep seeing these problems.”