Google tends to explain more often how do you work with different machine learning models to improve the performance of your applications. He did this recently explaining how radical Soli works and now explains how the new Google Duo quality improvements work.
He accomplished this with a new feature called WaveNetEQ, a comprehensive software based on DeepMind technology. the ability to complete lost packets in voice waves. Let's take a look at how Google accesses such a feature.
This is how Google improves Duo's audio quality with a display model
Google says when a call is transmitted over the Internet packages that have some quality issues. These problems are caused by excessive flexibility or network delays, and up to 8% of the total content can be lost due to those.
Of course you have encountered canned words or robots in a video call. The main reason is the loss of quality in the shipping and receiving process, if many are lost, the quality falls
To ensure that communication works well in real time, Google has created WaveNetEQ, a structured PLC (regulatory policy program) trained with a human-friendly voice database. What does the model do? The description is technical and sophisticated, so let's summarize it in a simple way
WaveNetEQ is a production model that allows you to synchronize voice waves even when waveforms are lost. You've certainly had the video call quality and heard the sound of robots or metal. This is because the lack of packets is high (there is plenty of latency) and the audio cannot be reproduced in quality.
With neural networks Google is able to provide signal continuity in real time, reducing quality loss. Basically, because of voice data, the model "guesses" what it is meant to say and ends the wave with missing pieces.
These improvements are now being implemented in the Google Duo on Google Pixel 4, though the company confirms that the model will finally come to other centers soon, so it will be an app development, not just devices.
More details | Google