- Led by Professor Jennifer Smith, University of Glasgow, in collaboration with University of California, Berkeley.
Automatic Speech Recognition (ASR), the conversion of spoken language into text by computers, has made significant advances in recent years. Systems such as AmazonTranscribe and WhisperX can now produce impressive speech-to-text transcriptions, ranging from highly specialised medical jargon to fairly rambling influencers talking about skincare on YouTube.
Despite these advances, ASR systems perform poorly with non-standard, often stigmatized dialects, further exacerbating existing inequalities in society. Here, the researchers bring expertise in sociolinguistics and speech technologies, comparing ASR performance in African American English and Scots, and examining how bias in these systems can be mitigated, to allow more inclusive AI technologies to be developed.
Image credit: Yutong Liu & Kingston School of Art / Talking to AI 2.0 / Licenced by CC-BY 4.0
