OpenAI Whisper speech recognition algorithm suffers from hallucinations

According to researchers, OpenAI’s Whisper speech recognition system sometimes suffers from hallucinations, i.e. is engaged in inventing facts. Software engineers, developers and scientists have expressed serious concerns that this feature of the AI ​​algorithm could cause real harm, since Whisper is already used in medical settings, among others.

Image source: Growtika / unsplash.com

The tendency of generative neural networks to invent facts when answering user questions has been discussed for a long time. However, it is strange to see this feature in the Whisper algorithm, which is designed for speech recognition. Researchers have found that a speech recognition algorithm can include anything in the generated text, from racist comments to fictitious medical procedures. This can cause real harm as Whisper has begun to be used in hospitals and other healthcare settings.

A University of Michigan researcher who studied transcripts of public meetings generated by Wisper found inaccuracies in the transcription of 8 out of 10 audio recordings. Another researcher examined more than 100 hours of audio transcribed by Whisper and found inaccuracies in more than half of them. Another engineer said he found inaccuracies in almost all of the 26,000 transcripts he created using Wisper.

An OpenAI representative said that the company is constantly working to improve the quality of its neural networks, including reducing the number of hallucinations. He also added that company policy prohibits the use of Whisper “in certain high-stakes decision-making contexts.” “We thank the researchers for sharing their results,” the OpenAI spokesperson added.