Algorithms for estimating reverberation characteristics for single channel dereverberation

Abstract: Speech enhancement systems achieving a joint suppression of reverberation and background noise can be used in digital hearing aids, voice controlled systems or hands-free telephony. Demanding requirements for computational complexity, signal delay and speech quality must be fulfilled in order to achieve a satisfactory performance. The speech quality depends on how accurate the reverberation characteristics such as the reverberation time or the spectral variance of the late reverberant speech are estimated. In this thesis, an efficient algorithm for a blind reverberation time estimation based on maximum likelihood approach is introduced. The new algorithm allows to estimate reverberation times from a much wider range with acceptable accuracy. Variance of the late reverberant speech is another important quantity in dereverberation systems. Two late reverberant spectral variance estimation methods are compared with regard to estimation accuracy and computational complexity. Finally, the performance of the considered speech enhancement system is analyzed with the improved reverberation time estimator.