Towards Single-Channel Speech Separation in Noise and Reverberation

Many speech technologies, such as automatic speech recognition and speaker identification, are conventionally designed to only work on single speech streams. As a result, these systems can suffer severely degraded performance in cases of overlapping speech, i.e. when two or more people are speaking at the same time. Speech separation systems aim to address this problem by taking a recording of a speech mixture and outputting a single recording for each speaker in the mixture, where the interfering speech has been removed. The advancements in speech technology provided by deep neural networks have extended to speech separation, resulting in the first effectively functional single-channel speech separation systems. As performance of these systems has improved, there has been a desire to extend their capabilities beyond the clean studio recordings using close-talking microphones that the technology was initially developed on. In this dissertation, we focus on the extension of these technologies to the noisy and reverberant conditions more representative of real-world applications. Contributions of this dissertation include producing and releasing new data appropriate for training and evaluation of single-channel speech separation techniques, performing benchmark experiments to establish the degradation of conventional methods in more realistic settings, theoretical analysis of the impact, and development of new techniques targeted at improving system performance in these adverse conditions