Auditory-based algorithms for sound segregation in multisource and reverberant environments
DeLiang Wang, Nicoleta Roman · 2005
At a cocktail party, we can selectively attend to a single voice and filter out all the other acoustical interferences. This perceptual ability has motivated the emergence of a new field of study known as computational auditory scene analysis (CASA) which aims to build speech separation systems that incorporate principles of auditory organization. This dissertation investigates four aspects of CASA processing: location-based speech segregation in multisource environments, binaural tracking of multiple moving sources, binaural sound segregation in reverberant environments, and monaural segregation of reverberant speech. The principal cues used by the auditory system to determine locations are the interaural time difference (ITD) and interaural intensity difference (IID) between the two ears. We observe that within a narrow frequency band, modifications to the relative strength of the target source with respect to the interference trigger systematic changes for ITD and IID. Moreover, for a fixed spatial configuration, this interaction produces a characteristic clustering in the binaural feature space. Consequently, we propose a supervised learning approach to estimate the ideal binary mask using the estimated binaural features. A systematic evaluation in terms of signal-to-noise ratio (SNR) as well as automatic speech recognition (ASR) scores shows that the resulting system produces masks very close to the ideal binary ones in anechoic conditions. Furthermore, the model produces large speech intelligibility improvements with normal listeners. While the above binaural systems perform optimally in anechoic conditions, reverberation affects the ITD and IID cues and therefore degrades their performance. For reverberant conditions, we propose a binaural segregation system that combines target cancellation through adaptive filtering and a binary decision rule to estimate the ideal binary mask. Specifically, we observe a correlation between the attenuation produced by the target cancellation stage and the relative strength between target and interference which is used subsequently to determine the target dominant T-F units. A major advantage of the proposed system is that, while requiring a fixed target location, it imposes no restrictions on the number, location or content of the interfering sources. An extensive comparison using SNR as well as ASR results shows that our system outperforms standard two-microphone beamforming approaches. (Abstract shortened by UMI.)