Denoising Speech Based on Deep Learning and Wavelet Decomposition

The work proposed a denoising speech method using deep learning. The predictor and target network signals were the amplitude spectra of the wavelet-decomposition vectors of the noisy audio signal and clean audio signal, respectively. The output of the network was the amplitude spectrum of the denoised signal. Besides, the regression network used the input of the predictor to minimize the mean square error between its output and input targets. The denoised wavelet-decomposition vector was transformed back to the time domain by the output amplitude spectrum and the phase of the wavelet-decomposition vector. Then, the denoised speech was obtained by the inverse wavelet transform. This method overcame the problem that the frequency and time resolution of the short-time Fourier transform could not be adjusted. The noise reduction effect in each frequency band was improved due to the gradual reduction of the noise energy in the wavelet-decomposition process. The experimental results showed that the method has a good denoising effect in the whole frequency band.