Blind dereverberation based on auto‐correlation functions of framewise time sequences of frequency components

Abstract

Proposed is a new blind dereverberation method based on auto‐correlation function of framewise time series of each frequency component. Inverse filtering of source‐microphone transfer functions is widely employed for suppressing the effects of reflected waves, but this method cannot be employed for cases where source‐microphone transfer functions are not available or for time‐variant cases. Several methods of spectral subtraction have been proposed to cope with these cases. However, most of them require transfer functions among sources and microphones. The proposed method, however, does not require transfer function inverse filters. Moreover, the method can handle frequency characteristics of sound paths. To realize dereverberation, it is necessary to estimate the delay time and frequency characteristics of reflection, though most conventional methods assume flat frequency characteristics. The proposed method estimates these parameters based on auto‐correlation function of each time series of spectral components in frequency spectra obtained every frame‐shift interval. The proposed method improves dereverberation performance, comparing with methods that assume flat frequency characteristics. The proposed method raises average segmental SNR by 3.4 dB and reduces reverberation time from 390 to 109 ms for a room employed for experiments. [Work supported by Knowledge Cluster Project, MEXT, and by Academic Frontier Project Doshisha University.]