Study shows voice deepfakes can fool speech-recognition devices

A team of researchers from the University of Chicagofound that voice-copying algorithms can deceive both speech-recognition devices and the people who listen to them.

The researchers tested two of the best-known deepfake models — SV2TTS and AutoVC. They assembled a dataset of 90 voice clips up to five minutes long from publicly available sources. They also recruited 14 volunteers who provided voice samples and access to their devices.

They then tested both systems using the open-source software Resemblyzer, which compares voice recordings and returns a similarity score between samples. They also conducted an experiment in which they attempted to gain access to voice-recognition-enabled devices using synthesized speech samples.

The researchers found that the algorithms managed to fool Resemblyzer in nearly half the cases. They also found that the deepfakes could deceive Microsoft Azure’s speech recognition service in about 30% of cases, and the Amazon Alexa assistant in 62% of cases.

Two hundred volunteers also listened to pairs of recordings and tried to determine whether the voices belonged to the same person. In most cases, the algorithms managed to deceive the participants, especially when the samples related to well-known individuals.

According to the researchers, voice-synthesis tools can cause serious damage in a variety of settings:

“They can bypass voice-authentication systems, automated telephone lines in banks, entry to messaging services such as WeChat. This could also lead to breaches of services that grant access to Internet of Things devices, for example Amazon Alexa or Google Home.”

The researchers added that voice deepfakes can also directly target end users, complementing traditional phishing scams with a familiar human voice.

In September, scammers used a deepfake of Oleg Tinkov to advertise a fake site ‘Tinkoff Investments’.

In mid-September researchers explained how to tell a fake from a real photo.

In April, an expert stated that the number of deepfakes online doubles every six months.

In March, American police arrested a woman for harassing teenagers using generated photos.

Subscribe to ForkLog news on Telegram: ForkLog AI — all the news from the world of AI!