Clean Speech Feature Estimation based on Soft Spectral Masking

In this paper, we first analyze the problems of speech and noise contamination process in noise-masking point of view, and propose a new approach to estimate degree of noise masking effect on clean speech distribution model based on sequential noise estimation. Sequential noise estimation is performed frame-by-frame using interacting multiple model (IMM) algorithm, so that real-time implementation is possible. After applying IMM algorithm, degree of noise masking effect named as noise masking probability(NMP) is calculated. Estimation of clean speech spectrum in noisy environments is performed by controlling the advantages of log spectrum domain and those of linear spectrum domain algorithm based on NMP. We have performed recognition experiments under noise conditions using the AURORA2 database which is developed for a standard reference of speech recognition performance. Simulation results show that this approach is effective when noise masking effect is dominated at low SNR.