Method for performing voice denoising by using single noisy voice sample

Abstract

The invention provides a method for voice denoising by using a single noisy voice sample. The method comprises the following steps: (1) for a clean voice signal, respectively superposing synthetic noise and different noise types in the real world to generate a noisy voice sample; (2) for a single noisy voice sample, generating a pair of voice training samples by using a voice down-sampler; (3) converting the trained input voice into a spectrogram, and then inputting the spectrogram into a denoising network for training, wherein the denoising network overlaps a plurality of two-stage Transformer modules between an encoder and a decoder of a ten-layer depth complex Unet; and (4) forming a loss function used for training by basic loss and regularization loss, wherein the basic loss is determined by network characteristics, and the regularization loss can prevent an excessive smoothness phenomenon from occurring in single sample denoising training. Compared with a traditional method for training by using clean voice and a pair of noisy voices, the method obtains better results on a plurality of evaluation indexes such as signal-to-noise ratio, voice quality perception evaluation, short-time objective intelligibility and the like.

Description

technical field [0001] The invention relates to a method for denoising speech by using a single noisy speech sample, and belongs to the fields of deep learning, speech denoising and speech enhancement. Background technique [0002] At present, electronic technology is widely used. As a typical non-stationary random signal, voice is the most commonly used medium for people to transmit information or communicate with each other. As voice services gradually emerge on smart terminals, people pay more and more attention to voice quality. Today, with the rapid development of informatization, voice signals will inevitably be interfered by various noises. People understand that it will also make it difficult for man-machine devices to obtain accurate information. Therefore, various speech denoising techniques have been developed and researched rapidly. In traditional research ideas, in order to achieve good speech noise reduction results, a large number of noisy speech samples and...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the deficiencies of the existing denoising network, the present invention integrates a complex module based on two-stage Tansformer in the complex encoder and decoder to learn the local and global context information output by the encoder to solve the long-term dependency problem of parallel computing, thereby Improving the Performance of Speech Denoising Networks

Images