Parallel spectral and cepstral modeling based speech enhancement using Hidden Markov Model

This paper is based on speech enhancement using Hidden Markov Model (HMM) in Mel-frequency domain. An inversion from the Mel-frequency domain to the spectral domain is required to estimate clean speech from a noisy speech signal. But it introduces distortion in spectrum. To reduce this effect, the Parallel Cepstral and Spectral (PCS) modeling is introduced. PCS method performs concurrent modeling in both magnitude spectral and cepstral domains. The performances of the PCS modeling are evaluated with different noise types at different SNR levels and the results are compared with conventional speech enhancement methods like MMSE, LSA and HNM-based speech enhancement. The experimental results show that PCS method is more efficient than other conventional methods.