Codebook-based speech enhancement with Bayesian LP parameters estimation

In this paper, we propose a codebook-based Bayesian linear predictive (LP) parameters estimation for speech enhancement, in which the LP parameters are estimated based on the current and past frames of noisy speech. First, by using hidden Markov model (HMM), we develop a new method to drive the speech presence probability (SPP) and speech absence probability (SAP). These two probabilities are the weighting coefficients for the estimated LP parameters corresponding to speech presence and speech absence states. Then we exploit the normalized cross-correction to adjust the transition probabilities between speech-presence and speech-absence states of HMM. The proposed adjustment method makes the SPP estimation more accurately. Finally, in order to suppress the noise between the harmonics of voiced speech, we employ the a posteriori SPP to modify the Wiener filter for enhancing the noisy speech. Our experiments demonstrate that the proposed method is superior to the reference methods.