Effects of reverberation on monaural speech separation and recognition

Abstract

Speech recognition is an active area of research with implementations spanning from commercial to medical applications such as hearing aids. Recognition of speech signals is studied for decades and a lot of progress has been shown in this area of research but there is still a lot of room for further research due to the adverse effects of the environmental conditions that contaminate clean speech signals. Such adverse conditions include noise and reverberation. Under such conditions, the recognition of automatic speech recognizers is subject to substantial degradation. Speech separation where the goal is to separate speech signals belonging to more than one talkers speaking at the same time is also an unsolved problem. It gets even harder as adverse environmental conditions are added to the scenario. This research is aimed at studying the effects of reverberation on monaural speech separation and recognition, and increasing the recognition performance of mixed speech signals. In the context of this research, recognition of mixed monaural speech signals with different reverberation levels is implemented and the effects of reverberation are evaluated. Performance increase in recognition is accomplished by training the system with moderately reverberated speech signals and then with speech signals that have predefined constant reverberation levels. For comparison purposes, effects of reverberation on speech recognition in the single-talker scenario are also examined along with the performance increase obtained by training the system with moderately reverberated signals.