Denoising Speech Based on Deep Learning and Wavelet Decomposition

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