Speech enhancement based on hidden Markov model with discrete cosine transform coefficients using Laplace and Gaussian distributions

This paper presents a novel HMM-based speech enhancement framework based on Laplace and Gaussian distributions in DCT domain. We propose analytical procedures for training clean speech and noise models with the aim of Baum's auxiliary function and present two MMSE estimators based on Gaussian-Gaussian (for clean speech and noise respectively) and Laplace-Gaussian combinations in the HMM framework. The performance evaluation is done using SNR and PESQ measures and the results of the proposed techniques are compared with AR-HMM approach. Higher SNR improvement is achieved for the proposed method in the Gaussian-Gaussian case in comparison with AR-HMM and Laplace-Gaussian techniques for both nonstationary and stationary noises. A similar result is obtained in term of PESQ in the presence of nonstationary noise types.