Google has launched out a latest technology in Duo calls to boost the audio quality when the service is incapable to sustain a secure link called WaveNetEQ. It relies on technology from the DeepMind division of Google that aims to replace audio jitter with artificial noise produced using machine learning, which sounds just like human speech.
If you’ve ever made an online call, chances are you’ve had audio jitter experience. This happens when the audio data packets sent as part of the call get lost along the way or either arrives late or in the wrong order. Google states that 99% of Duo calls experience packet loss: 20% of Duo calls mislay more than 3% of their audio and 10% loses over 8%.
Any calling software has to deal with this packet loss somehow, but Google says certain packet loss concealment (PLC) processes will struggle to fill gaps of 60ms or more without robotic or repetitive sounding. WaveNetEQ’s approach is based on neural network technology from DeepMind and has been trained in 48 different languages on data from over 100 speakers.
Here are a few Google audio samples which compare WaveNetEQ to NetEQ, a widely used PLC technology. Here’s how it looks when it is attempting to replace packet failure 60ms:
However, there is a cap to how much audio the machine can substitute. Google’s software is designed to substitute brief sounds, not entire words. So it’s fading out after 120ms, and generating silence. Google says it tested the framework to ensure that no major new sounds were added. Moreover, all authentications must also take place on-device, as Google Duo calls are encrypted by default from end to end.
It’s a sleek little piece of technology that can make calls that much easier to understand when they fail on the phone. Due to the December feature drop in the smartphone, the system is already available for Duo calls made on Pixel 4 phones, and Google says it’s rolling it out to other unspecified handsets.