Speech enhancement method using stacked multiscale modules

Abstract

The invention discloses an end-to-end speech enhancement method using stacked multiscale modules. The end-to-end speech enhancement method using the stacked multiscale modules comprises the followingsteps of S1, building a cascading end-to-end speech enhancement frame, and splicing the stacked multiscale modules into a network structure; S2, in the preprocessing stage, converting a time-domain signal into a two-dimensional characteristic; S3, utilizing a speech enhancement module for enhancing the two-dimensional characteristic; and S4, in the postprocessing stage, converting enhanced character representation into a one-dimensional time-domain signal through decoding synthesis. In order to further improve the performance of an algorithm, speech enhancement evaluation indicators STOI and SDR are integrated into a loss function through applying a multi-target joint optimization training strategy. The experiment shows that the method provided by the invention is capable of remarkably improving a speech enhancement effect and has better noise immunity under the conditions of unknown noise and low signal-to-noise ratio.

Description

technical field [0001] The invention belongs to the technical field of speech enhancement, in particular to an end-to-end speech enhancement method using stacked multi-scale modules. Background technique [0002] Speech enhancement refers to the task of removing or attenuating additional noise in noisy speech. It improves the overall perceptual quality of speech and speech intelligibility by suppressing and separating noise. In terms of robust speech recognition, hearing aid design, speaker verification, etc. Has a wide range of applications. Traditional speech enhancement methods include spectral subtraction, Wiener filtering, statistical model-based methods, and subspace-based methods. In the past few years, supervised speech enhancement methods based on deep learning have gradually become the main focus of scholars. research direction. [0003] Some scholars consider processing the time-domain signal of speech directly instead of relying on the frequency-domain represen...

AI Technical Summary

Problems solved by technology

[0004] These end-to-end methods directly map the one-dimensional time-domain waveform to the target speech. However, the time-domain waveform signal itself cannot show obvious characteristic structures, and it is difficult to directly model the time-domain signal. The difficulty of modeling will be further increased

Images