End-to-end voice enhancement method based on generation of countermeasure network

Abstract

The invention discloses an end-to-end voice enhancement method based on generation of a countermeasure network. The method comprises the following steps of directly inputting a noisy voice signal intoa pre-trained deep neural network for signal processing and outputting an enhanced voice signal, wherein the depth neural network is obtained through training by the following steps of S1, preliminarily training to generate a countermeasure network, wherein the generation of the countermeasure network comprises two deep neural networks: a generator G and a discriminator D; S2, performing knowledge distillation on the simulated noisy voice through a traditional statistical speech enhancement algorithm, and then training to generate a countermeasure network again; S3, performing fine adjustmenton the generator G obtained through training through the real noisy voice; and S4, outputting the generator G trained in the above step as a final deep neural network for voice enhancement processing.

Description

technical field [0001] The invention relates to the technical field of speech signal processing, in particular to an end-to-end speech enhancement method based on a generative confrontation network. Background technique [0002] Single-channel speech enhancement has been studied for decades, but it still faces great challenges in various application systems such as automatic speech recognition, hearing aids, and hands-free mobile communications. Traditional speech enhancement algorithms are usually based on statistical methods, including noise estimation and speech estimation. Since traditional speech enhancement algorithms are based on statistical model assumptions for speech signals and noise signals, their performance largely depends on the accuracy of noise estimation. [0003] In order to avoid the use of specific distortion criteria and model assumptions in the process of speech and noise signal processing, researchers have proposed a large number of data-driven speec...

AI Technical Summary

Problems solved by technology

Obviously, the computational cost of training multiple different models is too high; at the same time, the noisy speech in the actual environment cannot be obtained as pure speech as labeled data. Therefore, DNN often performs well in training data, while in some Real applications in specific environments have mediocre performance

Images