| In multimedia teleconferencing systems, an electronically steerable array of microphones has the potential to replace and outperform traditional head-mounted or hand-held and locally placed microphones as an input transducer system for obtaining high quality sound. However, faithful transduction of acoustic sources in enclosures is hampered by multipath transmission, attenuating interfering talkers, and ambient noise. For additive noise uncorrelated with the signal of interest, traditional speech enhancement methods work well. Another disturbing phenomenon, reverberation, is harder to attack because the effects of this distortion are convolutional and highly non-stationary. An effective algorithm must be not only capable of doing blind deconvolution for enhancing speech quality but also be computationally non-demanding to make it feasible for real-time implementations. This thesis concentrates on post-filtering deverberation methods incorporating the channel statistical properties into dereverberation techniques for estimating the anechoic speech. From this perspective, studying this problem has several advantages: the channel statistics are being taken into account, the sensor placement is incorporated source localization algorithm, and the Wiener postfilter accommodates multiple coherent source scenarios. |
