Neural network taught to simulate the interlocutor's speech during interruptions in audio calls

Due to an unstable connection during a call, the speech of the person on the other end of the line may be heard

intermittently. Google developers have solved this problem. They taught the neural network to imitate the interlocutor’s speech during interruptions in audio calls.

The fact is that when making video calls over the InternetDue to the signal passing through multiple networks, some audio packets may be lost. All services now use packet loss concealment (PLC) algorithms. However, if too many packets are lost and the delay increases to several tens of milliseconds, distortion may occur.

Developers from Google and DeepMind introducedA special algorithm that can create a replacement for lost packets. Based on the neural network algorithm. She analyzes the last fragments of speech, and then fills the pause with the synthesized voice of the interlocutor.

The technology has already been tested on Pixel 4 smartphones.