Echo cancellation method combining voiceprint recognition and deep learning

Abstract

The invention discloses an echo cancellation method combining voiceprint recognition and deep learning, and the method comprises the steps: 1) extracting acoustic features from a voice segment of a target speaker, and obtaining an embedded vector, i.e., a voiceprint vector, representing the information of the speaker through a pre-trained voiceprint encoder comprising a recurrent neural network model; and 2) extracting the acoustic features from the received signals, wherein the signals comprise near-end signals and far-end signals, estimating appearing echo signals, namely estimated echo signals through an AEC algorithm, and meanwhile obtaining near-end signals, namely linear filtering signals, containing residual echoes and eliminated in a linear mode. According to the echo cancellationmethod combining voiceprint recognition and deep learning provided by the invention, echo cancellation can be realized under the conditions of double speaking, non-occurrence distortion and the like,and the echo cancellation effect is greatly improved even under the condition that a far-end signal contains human voice.

Description

technical field [0001] The present invention relates to the technical field of echo cancellation, and more particularly, to an echo cancellation method combining voiceprint recognition and deep learning. Background technique [0002] When the smart speaker is playing an audio program, the signal collected by the microphone includes both the signal of the target speaker (near-end signal) and the audio signal played by the speaker (far-end echo). In this case, the speech recognition module is almost can not work. It is usually necessary to introduce an echo cancellation method to eliminate the echo and protect the near-end voice signal from loss. The current mainstream echo cancellation method is to use the reference signal collected by the hardware loop to estimate the echo signal existing in the microphone signal, subtract the estimated echo signal from the original signal to obtain the estimated near-end signal, and finally pass a post- Non-linear filter to suppress resid...

AI Technical Summary

Problems solved by technology

[0004] In order to solve the technical problem that the effect of echo cancellation is not good when the far-end signal contains human voice, the present disclosure provides an echo cancellation method that combines voiceprint recognition and deep learning

Images