Bias in Speech Recognition

Author: Jamie Brandon

Publisher: Springer Link

Publication Year: 2020

Summary: The following article discusses how speech recognition algorithms are less effective when it comes to Black people. An algorithm was given a matched subset of identical short phrases spoken by Black and white speakers. The algorithm produced a higher word error rate for Black speakers than for white speakers. Algorithms that trend with demographic data will tend to suffer more in capturing the effects of underrepresented datasets. In this case, Black people were not represented accurately in the dataset, which led to results. Artificial Intelligence, like Siri and Alexa, sometimes does poorly in identifying phrases from black speakers. When dealing with the demographic dataset, underrepresented proportions must be accurately measured.