Stereo echo cancellation method and system based on neural network

Abstract

The invention discloses a stereo echo cancellation method and system based on a neural network, and the method comprises the steps: enabling the system to input three audio signals: a microphone signal, a left channel reference signal played by a stereo loudspeaker, and a right channel reference signal played by the stereo loudspeaker; performing short-time Fourier transform on the three paths ofaudio signals to obtain three paths of frequency domain signals, compressing the three paths of frequency domain signals to form frequency band energy information, taking the three paths of frequencyband energy information as input of a neural network model, and outputting an ideal probability masking IRM value through the trained neural network model, wherein the neural network model takes a setIRM value as a training target; obtaining frequency domain information after echo cancellation according to the IRM value obtained through calculation and the frequency band information energy valuecorresponding to the microphone signal; and carrying out inverse short-time Fourier transform on the frequency domain information to obtain time domain information, and completing echo cancellation. According to the invention, an artificial intelligence-deep learning method is adopted to replace a traditional complex stereo echo cancellation algorithm, and stereo echo cancellation in voice communication is realized.

Description

technical field [0001] The invention relates to the technical field of real-time communication and speech recognition, in particular to a neural network-based stereo echo cancellation method and system. Background technique [0002] In voice communication, echo is an important factor affecting call quality. The echo phenomenon refers to the voice that the speaker sends to other people through the communication device and returns to the speaker's own earpiece. The echo is a serious interference to the speaker and must be eliminated. When the playback volume of the near-end user is relatively high and the distance between the recording device and the playback device is relatively close, the echo is particularly obvious. The echo is affected by the environment of the near-end user, and may generate multiple echoes, including direct echoes and reflected echoes. The paths of each echo are different, and the delay is also different, so it is difficult to eliminate. figure 1 The...

AI Technical Summary

Problems solved by technology

However, the existing decorrelation algorithms are not very effective, and the voice quality is also degraded while decorrelating the signals; although many complex adaptive filtering algorithms have been developed, due to the distance between the far-end signals Strong correlation, slow convergence speed of adaptive filtering algorithm, sensitive to changes in the echo path between the far-end room and the near-end room, and the algorithm is not robust

Images