Media reported that Google has launched a sound improvement program called WaveNetEQ to ensure audio quality for Duo’s call service in an unstable connection. It is understood to be based on technology from Google’s DeepMind division, which aims to use artificial noise to suppress audio jitter. This noise sounds like a human being talking, but it’s actually generated by machine learning.
If you’ve ever made a network call, you’ve definitely experienced more or less audio jitter due to unstable connections.
This can happen when some of the audio packets during a call are lost, delayed too high, or in the wrong order.
Google says 99 percent of Duo calls are lost, with 20 percent losing more than 3 percent and another 10 percent losing more than 8 percent.
Every communications app has a way of dealing with packet drops, but Google says the traditional packet-loss (PLC) process could lead to vacancies of up to 60ms.
By contrast, WaveNetEQ solutions based on DeepMind neural network technology have been trained on more than 100 speaker data in 48 different languages.
To add to its persuasiveness, Google has also released some of the comparison audio of WaveNetEQ with common PLC technologies. Of course, there are limits to how much audio new technology can convert.
Google says WaveNetEQ is designed to replace shorter audio intervals, rather than the entire single missing during a live call. So after 120ms, the system fades out and mutes.
Through the assessment, Google ensured that WaveNetEQ would not introduce the accent of any new operations, and that all processing could be done on the user’s device side, after all, Google Duo turned on end-to-end encryption by default.
As soon as the network connection becomes smooth, live audio is restored and WaveNetEQ seamlessly fades out of the device. The technology has been tested on the Pixel 4 smart computer and will be rolled out to more devices.