Neural dual particle filter and its application in speech enhancement

Traditional speech enhancement techniques are commonly spectral methods, which frequently result in audible distortion of the signal. In this paper, a neural network based time-domain method called dual particle filter (dual PF) is proposed for speech enhancement, which consists of two PFs run concurrently. At each time-step, two PFs estimate both the state and model from only noisy observations respectively. We apply this method on the speech enhancement in the presence of both white (stationary and nonstationary) and colored noise. The experiments show that the approach performs significantly better than the traditional techniques on the reduction of white noise, and performs well in the presence of stationary colored as well.