Speech recognition and racial bias

Studies have recently been shown that facial recognition systems develop racial bias because they are trained on mainly Caucasian face image sets.

As speech technology experts, we wondered if there was this type of bias in voice search engines as well.

Now a new American study has found that even the speech recognition systems of large tech companies have this type of bias.

The researchers subjected nearly 20 hours of interviews to voice recognition systems by transcribing 42 interviews from whites and 73 from blacks. The average error rate for white respondents was 19%, while for black respondents it was 35%.

We, therefore, propose what happened for the facial recognition in which people with darker skin tones are erroneously identified in a higher percentage.

This is since the data sets on which the sampling is carried out are mainly provided by whites.

The Microsoft system has achieved the best overall result, with an error rate of 15% for white recordings and 27% for black recordings. Apple posted the worst results, with an error rate of 45% for blacks and 23% for whites.

#Voicebot #VocalMultichannel #chatbot