Complex field speech enhancement method and system based on generative adversarial network and medium

Abstract

The invention discloses a complex field speech enhancement method and system based on a generative adversarial network and a medium. The complex field speech enhancement method comprises the implementation steps that a noisy speech is acquired; the speech is expressed with Cartesian coordinates after being subjected to Fourier transformation to obtain a noisy real number spectrum and imaginary number spectrum; the noisy real number spectrum and imaginary number spectrum are input into a generator, which completes training in advance, of the generative adversarial network, and a denoised real number spectrum and imaginary number spectrum of a pure speech are obtained; and the real number spectrum and imaginary number spectrum of the pure speech are made into a clean speech based on inverseFourier transformation. Through the method, noise can be better removed from a speech signal to generate the clean speech, the problem that it is difficult to predict a phase is effectively solved, the auditory effect of the enhanced speech can be effectively improved, and the speech recognition accuracy rate of a speech recognition system in a noise environment can be effectively increased.

Description

technical field [0001] The present invention relates to a speech noise reduction and enhancement technology based on a generative confrontation network, in particular to a method, system and medium for speech enhancement in the complex domain based on a generative confrontation network. The signal is enhanced to facilitate the research of related downstream tasks such as speech recognition. Background technique [0002] Speech Enhancement (Speech Enhancement, SE) refers to removing the noise z from the noisy speech y, so as to separate the pure speech x, that is, x=y-z. Removing noise from a mixed speech signal is one of the most challenging tasks in speech signal processing. Traditional speech enhancement algorithms include spectral subtraction, subspace method and Wiener filtering method. In recent years, deep learning-based speech enhancement techniques have greatly improved the quality of denoised speech. [0003] In a general speech signal processing method, the speec...

AI Technical Summary

Problems solved by technology

The generator and the discriminator are trained against each other. The samples generated by the generator try to obey the real distribution to confuse the discriminator so that the discriminator can judge it as true, and the discriminator tries to separate the real samples from the generated samples. In this continuous game In the process until the Nash equilibrium is reached, the samples generated by the generator are very close to the real samples at this time, and the discriminator cannot judge whether the generated samples are real or generated

Images