Two stage blind dereverberation based on stochastic models of speech and reverberation

Kavruk, Mehmet
Distant speech processing is popular nowadays due to wide use of the hands-free communication with smart devices. The quality of microphone signals in an enclosed area is degraded by environmental noise and reverberation in distant speech communication. Although there are powerful denoising algorithms in the literature, there is no robust dereverberation method which works independent of recording conditions. This work proposes a statistical model based blind dereverberation algorithm which suppresses reverberation part without causing serious degradation in the source signal in different speaker to microphone configurations. The proposed algorithm successively uses minimum variance distortionless response (MVDR) and linear prediction methods. The parameters of the MVDR algorithm are estimated using the statistical nature of reverberation. The linear prediction algorithm is applied to the output of MVDR in order to handle residual reverberation. The dereverberation filter in this stage is generated using the statistical models of speech and reverberation. None of the algorithms require any deterministic prior knowledge about the system due to the used statistical models. The experimental results demonstrate that the proposed algorithm suppresses reverberation in the distant recordings without degradation on the source signal with respect to the objective quality measures under different conditions.
Citation Formats
M. Kavruk, “Two stage blind dereverberation based on stochastic models of speech and reverberation,” M.S. - Master of Science, Middle East Technical University, 2019.