Generative adversarial network speech enhancement method based on sparse continuous constraint

Abstract

The invention discloses a generative adversarial network speech enhancement method based on sparse continuous constraint. The method comprises the following steps of: 1) collecting and classifying data; 2) carrying out speech framing and windowing; 3) carrying out amplitude compression; 4) inputting sparsity constraint-based generative adversarial network training; 5) carrying out amplitude decompression; and 6) synthesizing an enhanced speech. The method has the advantages that distribution of clean voice samples is finally obtained through adversarial learning between the generative model and the discrimination model in the generative adversarial network; there is no assumption for statistical distribution of voice or noise; and sparsity and continuity constraints are added to a loss function of the generator, so that the obtained sparse matrix can better conform to speech spectrum distribution. According to the method, the problem that the voice and noise signal distribution is difficult to estimate is ingeniously solved, the voice intelligibility is improved, and the enhanced speech more conforming to the pure voice spectrum distribution is obtained.

Description

technical field [0001] The invention relates to the technical field of voice processing, in particular to a method for enhancing voice over a generative confrontation network based on sparse continuous constraints. Background technique [0002] As the main medium of human communication, voice has already played an important role in mobile communication, voice assistant and other fields. Against the background of artificial intelligence technology and Internet of Things technology in the ascendant, the wide application of speech recognition, voiceprint recognition and other technologies also put forward higher requirements for the quality of speech signals. However, in actual speech acquisition and dialogue communication scenarios, speech signals are often interfered by various noises. Speech enhancement is an effective technique to solve noise pollution. [0003] There are four main traditional speech enhancement methods: (1) spectral subtraction, which uses the short-term...

AI Technical Summary

Problems solved by technology

[0004] The problem to be solved by the present invention is to provide a generative adversarial network speech enhancement method based on sparse continuous constraints to solve the problem of "music noise", non-stationary noise interference, and real speech distribution in the existing methods in the environment of low signal-to-noise ratio. inestimable problem

Images