Machine learning is essential for Google, as it improves the performance of its applications and services. Recently, they told us about the way the Pixel 4’s sensor worked . The brand does it now with the improvements in the audio of the calls in Google Duo, which have a new model that helps improve them, as they have explained in their blog.
This new model is called WaveNetEQ and it is a generative system based on Deep Mind technology . It stands out for having the ability to complete lost packets in voice waves, which is why it helps the audio quality of calls in Google Duo improve dramatically.
Improving calls on Google Duo
When a call is transmitted over the Internet, the packets have certain quality related issues. They can be caused by network delays or excessive fluctuations, causing a large amount of content to be lost. According to Google, up to 8% of the content is lost in this case. This is what has forced the company to take action.
So they have created WaveNetEQ , which is a PLC (programmable logic controller) system that has been trained with a voice database. This generative model allows to synthesize voice waves, even in those cases in which some parts of it have been lost. Avoiding said loss mentioned above.
Thanks to the use of neural networks, what Google does is that the signal has continuity in real time. The most direct consequence of this is that the quality is maintained, such loss of quality is minimized . This model will “guess” what it is intended to say and will complete this wave with those missing fragments, those that have been lost.
In this way, such calls in an application like Duo will maintain good quality, with good audio. These interruptions or loss of quality are avoided. Training the model has not been easy, but it is in a state where it already has the ability to predict what will be said.
Google Duo in the Pixel 4 are already starting to apply these improvements, so users will be able to enjoy better audio and less loss of quality in their calls. It is expected that this WaveNetEQ model will reach more devices, since it is an improvement in the application, some models will not be limited only. Unfortunately, the dates for this are currently unknown.