Statistical models for speech dereverberation

This paper discusses a statistical-model-based approach to speech dereverberation. With this approach, we first define parametric statistical models of probability density functions (pdfs) for a clean speech signal and a room transmission channel, then estimate the model parameters, and finally recover the clean speech signal by using the pdfs with the estimated parameter values. The key to the success of this approach lies in the definition of the models of the clean speech signal and room transmission channel pdfs. This paper presents several statistical models (including newly proposed ones) and compares them in a large-scale experiment. As regards the room transmission channel pdf, an autoregressive (AR) model, an autoregressive power spectral density (ARPSD) model, and a moving-average power spectral density (MAPSD) model are considered. A clean speech signal pdf model is selected according to the room transmission channel pdf model. The AR model exhibited the highest dereverberation accuracy when a reverberant speech signal of 2 sec or longer was available while the other two models outperformed the AR model when only a 1-sec reverberant speech signal was available.