A gain-adaptive parallel HMM for speech enhancement

This document is an example of what your final camera-ready manuscript to APSIPA ASC 2015 should look like. Authors are asked to conform to the directions reported in this document. The key problem in Hidden Markov model (HMM)-based speech enhancement is how to obtain an appropriate weighted Wiener filter (WWF). This paper presents a gain-adaptive parallel HMM (PHMM) for speech enhancement based on Mel-frequency spectral (MFS) features and linear prediction (LP) coefficients of speech and noise, and a WWF modified by the speech-presence probability (SPP) is developed. MFS-HMM (i.e., the HMM in MFS domain) and LP-HMM (i.e., the HMM in LP domain) constitute the proposed PHMM, which is obtained by the parallel training method. The forward probabilities in all mixtures of states for noisy speech MFS-HMM are calculated as the weighting factors of WWF. In order to obtain more accurate noisy speech MFS-HMM and solve the mismatching problem of spectral energy between the training and test signals in MFS domain, we introduce two gain factors to adaptively adjust the spectral energy of speech and noise, respectively. The gain factors are estimated online by the expectation maximization (EM) algorithm. The pre-trained LP-HMMs of speech and noise only contain the spectral shapes of speech and noise, so the EM algorithm is also employed to estimate LP spectral gains for constructing Wiener filters in the proposed WWF. The evaluation results confirm the superiority of the proposed method.