HMM-based speech enhancement using vector Taylor series and parallel modeling in Mel-frequency domain
Zhenzhen Gao, Changchun Bao, Feng Bao, Maoshen Jia · 2014
Speech enhancement based on hidden Markov model (HMM) and the minimum mean square error (MMSE) criterion in Mel-frequency domain is generally considered as a weighted-sum filtering of the noisy speech. The weights of filters are often estimated by the HMM of noisy speech, and the estimation of filters usually requires an inverse operation from the Mel-frequency to the spectral domain which often causes spectral distortion. In order to obtain a more accurate HMM of noisy speech, the vector Taylor series (VTS) is used to estimated the mean vectors and covariance matrices of HMM for noisy speech. To reduce the distortion derived from inversion operation, a parallel Mel-frequency and log-magnitude (PMLM) modeling approach is proposed. In PMLM, a simultaneous modeling in both Mel-frequency domain and log-magnitude (LOG-MAG) domain is performed to train the HMMs of the clean speech and noise. Experimental results show that, in comparison with the reference methods, the proposed method can get better performance for different noise environments and input SNRs.