A Double-Noise Speech Enhancement Method Using Multiple Modules to Suppress Different Kinds of Noise

Abstract

The invention discloses a dual-noise speech enhancement method for multi-module suppression of different types of noise, which includes the following steps: S1: performing stage-by-stage modeling on various types of noise, and for the input noisy speech, through noise suppression at each stage The module extracts and filters one or more noise features; among them, the loss function of each noise suppression module is different; S2: the amplitude spectrum of the suppressed part of the noise in the process and the amplitude spectrum of the original noisy speech are spliced ​​and input into the final neural network network; the present invention proposes a multi-module staged dual-noise speech enhancement method that suppresses different types of noise. performance, and then integrate the enhanced results into the next stage, it uses the neural network to learn the mapping from the noisy magnitude spectrum to the purer magnitude spectrum at each stage, and refines the features to obtain a purer magnitude spectrum.

Description

technical field [0001] The invention belongs to the technical field of speech enhancement, in particular to a dual-noise speech enhancement method in which multiple modules suppress different types of noise. Background technique [0002] Speech enhancement algorithms are an important speech processing technology that power speech recognition systems, hearing aids, and military bugging devices. At present, the accuracy rate of speech recognition algorithms has reached a relatively high level, even surpassing skilled dictation transcriptionists in some public data sets. However, due to the existence of noise or reverberation interference, the speech recognition algorithm can only achieve the desired effect after speech enhancement. The current speech enhancement algorithms only perform well on noisy speech containing a single noise with a high signal-to-noise ratio. In real scenes such as meeting environment, battlefield environment and street environment, there will be many...

AI Technical Summary

Problems solved by technology

For dual-noise scenes, general algorithms are not easy to show good generalization

In a low SNR environment, the performance of the speech enhancement algorithm will deteriorate significantly due to the dominant noise energy in the audio.

Images