Training method and device of speech enhancement model a well as speech enhancement method and device

Abstract

The invention relates to a training method and device of a speech enhancement model as well as a speech enhancement method and device. The training method comprises the steps that: the feature vectors of noisy speech samples and first pure speech samples of a plurality of speakers are obtained, wherein the noisy voice sample of each speaker is obtained by adding noise data to a second pure speech sample corresponding to the speaker; the amplitude spectra of the noisy speech samples are input into a speech enhancement network to obtain an estimated first mask ratio; the estimated first mask ratio and the feature vector are input into an attention mechanism network to obtain an estimated second mask ratio; an estimated amplitude spectrum is determined according to the estimated second mask ratio and the amplitude spectra, and a loss function of the speech enhancement model is determined according to the estimated amplitude spectrum and the amplitude spectra of the second pure speech samples; and the speech enhancement model is trained by adjusting parameters of the speech enhancement network and the attention mechanism network according to the loss function.

Description

technical field [0001] The present disclosure relates to the field of audio technology, and more specifically, to a method and device for training a speech enhancement model, and a method and device for speech enhancement. Background technique [0002] Noisy environment will affect the effect of people's voice communication. In the current mainstream communication software, different voice enhancement algorithms are usually used to process the noise-containing frequency during the call. The traditional method can realize the processing of steady-state noise. Advantages It is low computational complexity. Deep learning methods are usually used to remove transient noise. The effect is better than traditional methods, but the computational complexity is high. [0003] Noisy speech usually contains background noise or the voices of other speakers. In order to improve communication efficiency, it is necessary to obtain the pure speech of a specific speaker. Conventional speech en...

AI Technical Summary

Problems solved by technology

[0002] Noisy environment will affect the effect of people in voice communication. In the current mainstream communication software, different voice enhancement algorithms are usually used to process the noise-containing frequency during the call. The traditional method can realize the processing of steady-state noise. Advantages The computational complexity is low. The deep learning method is usually used to remove transient noise. The effect is better than the traditional method, but the computational complexity is high.

[0003] Noisy speech usually contains background noise or the voices of other speakers. In order to improve communication efficiency, it is necessary to obtain the pure speech of a specific speaker. Conventional speech enhancement can remove background noise and separate the voices of each speaker. But still facing the problem of sorting the speakers, I don’t know which speaker’s voice should be output, so the effect of speech enhancement for specific speakers is general

Images