High-Frequency Component Restoration for Kalman Filter Based Speech Enhancement
Hongjiang Yu, Wei‐Ping Zhu, Benoı̂t Champagne · 2020
In this paper, we present a deep neural network (DNN) based algorithm to restore the high-frequency (HF) component of the enhanced speech processed by Kalman filtering, where the DNN is applied for estimating the magnitude of HF component from the low-frequency (LF) counterpart. The complete HF component is then computed with the estimated magnitude given by the DNN and the phase of the Kalman filtered speech. By incorporating our restoration algorithm into Kalman filter based speech enhancement method, our new speech enhancement system is able to recover the HF component with better perceptual quality and less distortion. Experimental results demonstrate that the proposed method outperforms the state-of-the-art Kalman filter based method in terms of both speech quality and intelligibility.