Source Separation and Speech Dereverberation

11.4 Summary

Although humans can easily distinguish one voice amid a clutter of distracting speech-like interference and background noise in reverberant environments, it is not easy to decipher this ability and it is even more challenging to develop effective source separation and speech dereverberation algorithms to let machines mimic this processing in our brains. After decades of continuous research surrounding this phenomena by psychoacousticians, neural scientists, and signal processing engineers, fascinating, if not overwhelming, advancement has been achieved.

This chapter studied the state-of-the-art techniques of source separation and speech dereverberation. We began with an overview of the cocktail party effect and explained what implications we can derive for developing source separation algorithms. Then we had a survey of microphone arrays beamforming. We analyzed the direction-oriented signal model, illustrated various beamformer structures, commented on the difference between narrow band and broadband beamforming techniques, and discussed the popular linearly constrained minimum-variance (LCMV) and generalized sidelobe cancellation (GSC) algorithms. We have also reviewed the emerging independent component analysis (ICA) methods for blind source separation (BSS). Both cutting-edge high-order-statistics ICA and second-order-statistics BSS algorithms were developed. In the final section, we have presented a synergistic solution to source separation and speech dereverberation based on blind identification of acoustic MIMO systems. Its usefulness was justified by realistic experiments.