Google has confirmed contractors around the world listen to samples of audio recorded by the virtual Google Assistant on smart speakers and Android devices. The confirmation comes after a rogue Dutch contractor reportedly shared more than 1,000 audio clips with Belgian public broadcaster VRT NWS, which notes 153 conversations were clearly not directed at a Google device. The devices are typically activated with the tap of a button or hotwords like "OK, Google," but VRT News found that the recording can start when something that loosely resembles a hotword is said. Per USA Today, the conversations included fights between couples, explicit bedroom talk, confidential business calls, and discussions with children. "In these recordings we could clearly hear addresses and other sensitive information,” reads the VRT NWS report. "This made it easy for us to find the people involved and confront them with the audio recordings."
In a Thursday blog post, Google—which is investigating the leak—said language experts listen to and transcribe 0.2% of its audio clips. "This is a critical part of the process of building speech technology, and is necessary to creating products like the Google Assistant," the company said, noting background noise may be "interpreted to be the hotword" in rare cases. Google's privacy policy also notes personal information is provided to "affiliates and other trusted businesses or persons" to be processed. An expert on voice technology notes Amazon, Microsoft, Samsung, and other companies offering virtual assistants likely do the same thing, per the Wall Street Journal. "Anything with speech recognition, you generally have humans at one point listening and annotating to sort out what types of errors are occurring," he says. (More Google stories.)