A parallel cepstral and spectral modeling for HMM-based speech enhancement

An HMM-based speech enhancement in Mel-frequency domain is introduced and improved. It is shown that hidden Markov modeling in the Mel-frequency domain is beneficial due to its effective representation of the speech spectrum; however, speech enhancement in this domain requires an inversion from the Mel-frequency to the spectral domain which introduces distortion artifacts for spectrum estimation. To reduce the distortion effects of the inversion and employ the advantages of robustness modeling in the Mel-frequency domain, a parallel cepstral and spectral (PCS) modeling is proposed. In PCS, a concurrent modeling in both cepstral and spectral domains is performed. The performances of the speech enhancement system using Mel-frequency spectral (MFS) and Mel-frequency cepstral (MFC) features and using the PCS modeling are evaluated on various corrupting noise types with different SNR levels. The results confirm the superiority of the proposed methods particularly in dealing with non-stationary noises.