Multiscale collaborative speech denoising based on deep stacking network

A growing number of noise reduction algorithms based on supervised learning have begun to emerge in recent years and show great promise. In this study, we focus on the problem of speech denoising at very low signal-to-noise ratio (SNR) conditions using artificial neural networks. The overall objective is to increase speech intelligibility in the presence of noise. Inspired by multitask learning (MTL), a novel framework based on deep stacking network (DSN) is proposed to do speech denoising at three different time-frequency scales simultaneously and collaboratively. Experiment results show that our algorithm outperforms a state-of-the-art method that is based on traditional deep neural network (DNN).