Google is hoping to end low quality video calls by deploying artificial intelligence to “fill in” audio gaps caused by bad connections.
WaveNetEQ works by using a library of speech data to realistically continue short segments of conversations.
The AI is trained to produce mostly syllable sounds, and can fill gaps of up to 120 milliseconds.
It comes as the use of video calls has become increasingly important during the corornavirus crisis.
When making a call over the internet, data is split into small chunks called packets.
A poor connection can mean these packets reach the other party in the wrong order and at the wrong time, or cause them to be lost entirely. This can result in a significant decline in call quality.
Google says 99% of calls made using its Duo app experience some form of audio-related issue. Of these calls, 20% lose more than 3% of their total audio, while 10% lose almost a tenth.
WaveNetHQ works by specifically creating speech data to fill the gaps made by drops in audio.
Data-sharing
The AI has been trained using the voices of 100 individuals in 48 languages to enable it to learn the general characteristics of a human voice, regardless of dialect.
Douglas Crawford, cyber security researcher at ProPrivacy, says that Duo’s end-to-end encryption should help alleviate any concerns about data-sharing.
“As calls on the platform are secured using end-to-end encryption, outsourcing AI-processing of missing packets in order to reduce audio jitters was simply not an option for developers,” he told the BBC.
“Google solved this by performing all the processing on your device so that no data is ever transmitted to a third party.
The system is currently available on Google’s Pixel 4 smartphone – the company says it plans to expand to more Android devices later this year.
In 2018, Google divided critics when it unveiled artificial intelligence software that books appointments over the phone on behalf of users by making realistic voice-based calls. However, the feature is currently only available in the US.
