Acoustic source separation based on target equalization-cancellation

Normal-hearing listeners are good at focusing on the target talker while ignoring the interferers in a multi-talker environment. Therefore, efforts have been devoted to build psychoacoustic models to understand binaural processing in multi-talker environments and to develop bio-inspired source separation algorithms for hearing-assistive devices. This thesis presents a target-Equalization-Cancellation (target-EC) approach to the source separation problem. The idea of the target-EC approach is to use the energy change before and after cancelling the target to estimate a time-frequency (T-F) mask in which each entry estimates the strength of target signal in the original mixture. Once the mask is calculated, it is applied to the original mixture to preserve the target-dominant T-F units and to suppress the interferer-dominant T-F units. On the psychoacoustic modeling side, when the output of the target-EC approach is evaluated with the Coherence-based Speech Intelligibility Index (CSII), the predicted binaural advantage closely matches the pattern of the measured data. On the application side, the performance of the target-EC source separation algorithm was evaluated by psychoacoustic measurements using both a closed-set speech corpus and an open-set speech corpus, and it was shown that the target-EC cue is a better cue for source separation than the interaural difference cues