World

Speech recognition company Speechmatics has defeated tech giants by reducing bias

A young woman using voice recognition software on her smartphone.

Luis Alvarez | DigitalVision | Getty Images

London — A speech recognition company in the United Kingdom states that its software outperformed major technology companies such as: Google When Amazon In understanding the black voice.

Cambridge-based Speechmatics in the United Kingdom said Tuesday that the overall accuracy rate for African-American voices in the system was 83%.

It’s higher Microsoft (73%), Amazon (69%), Google (69%), IBM (62%) and Apple According to (55%) research Published by Stanford University in 2020, we compared the results of how well the speech recognition programs of leading technology companies understand African Americans.

According to researchers at Stanford University, the Amazon, IBM, Google, Microsoft, and Apple systems made almost twice as many errors as whites in interpreting words spoken by African Americans.

According to Speechmatics, the system had a 17% chance of misidentifying a word from a black voice, compared to 31% on Google and Amazon.

Alison Kennecke, lead author of research at Stanford University, said:

AI bias

With the widespread use of virtual assistants in smart devices such as phones and speakers, voice recognition technology is rapidly becoming an integral part of everyday life.

Apple has pioneered the use of voice-launching software on mobile devices with the digital assistant Siri. Amazon, on the other hand, was one of the first companies to bring speech recognition to the home using Echo speakers and Alexa assistants.

Researchers are becoming more and more concerned about the bias of algorithms that enhance these speech recognition services. Specifically, experts say that many speech recognition programs are less effective because they are trained on a limited set of datasets.

According to Will Williams, Vice President of Machine Learning at Speechmatics, speech recognition applications have few problems with transcription, but white, male, and East Coast news presenters said in an underrated voice, “the same level of accuracy. Do not have”.

“Like all this, it’s about the quality of the data in the training set,” AI ethics researcher Stephanie Hair told CNBC. “Speech recognition technology has long had racial prejudice, gender prejudice, and regional accent prejudice.”

“This technology still doesn’t work the same for everyone,” Hair added. “Ultimately, it can be sophisticated.”

Unlabeled data

Speech recognition company Speechmatics has defeated tech giants by reducing bias

Source link Speech recognition company Speechmatics has defeated tech giants by reducing bias

Back to top button