Reconstructing missing speech spectral components using both temporal and statistical correlations

This paper presents a new method for reconstructing unreliable spectral component which uses statistical distributions of former and later reliable frames and reliable components of current frame. In this technique, first, a HMM is used to model the temporal variation of clean speech signal. Then using this model and according to probabilities of occurring noisy component at each states, a distribution for noisy components is estimated. Finally, by applying MAP estimation on mentioned distribution final estimation of this unreliable component is obtained. The proposed method has been compared to a recent missing feature method which is based on clustering feature vectors and exhibits a significant enhancement in two different noisy environments.