Speech enhancement model training method, speech enhancement model recognition method, electronic equipment and storage medium

Abstract

The invention discloses a multi-task fusion speech enhancement model training method. The method comprises the following steps: acquiring an echo signal source, clean speech and real noise; and obtaining reverberation voice and noisy signals. And obtaining a training label of the to-be-trained neural network. And training to obtain a speech enhancement model. According to the embodiment of the invention, the nonlinear distortion characteristics introduced by a neural network learning hardware system can improve the suppression degree of residual nonlinear echoes. And better reverberation removal and noise reduction effects can be obtained by utilizing the advantages that the neural network is superior to the traditional linear prediction and noise estimation algorithm. Multiple speech enhancement tasks can be completed by a single model, algorithm integration is simple and efficient, and maintenance cost is low.

Description

technical field [0001] The invention belongs to the technical field of speech processing and recognition, and in particular relates to training of a speech enhancement model, a recognition method, electronic equipment and a storage medium. Background technique [0002] In the prior art, echo cancellation refers to estimating a transfer path function by using a microphone and a reference sound (echo signal source), and then subtracting the estimated echo from the microphone signal to suppress the echo. De-reverberation refers to the use of linear prediction to estimate the late reverberation, and then subtract the late reverberation signal from the original signal to obtain a mixed signal of the direct sound and the early reverberation. Environmental noise suppression refers to the use of noise estimation algorithms to estimate the noise and then remove it. [0003] Echo cancellation cannot estimate the nonlinear distortion of a hardware system (such as a loudspeaker), so th...

AI Technical Summary

Problems solved by technology

In the reverberation speech signal observation model, it is assumed that the observed signal does not contain noise, so when the observed signal contains noise, the noise may cause the filter performance to deteriorate

[0006] Traditional noise estimation algorithms, such as minimum tracking algorithm, time recursive averaging algorithm, histogram-based algorithm, etc., are all based on the assumption that the noise in the analysis time period is more stable than the speech, that is, the noise change rate is slower than the speech, so it cannot be accurate. and estimate abrupt non-stationary noise in time

In addition, there are certain differences in the mathematical models, assumptions, and optimization goals of these different algorithm modules, so it is impossible to integrate these algorithms into a unified algorithm

Images