Blind separation of speech sources in multichannel compressed sensing

Abstract

This paper presents a novel framework for separating and reconstructing multichannel speech sources from compressively sensed linear mixtures simultaneously. The conventional approaches for blind speech separation are almost based on the Nyquist sampling theory. We proposed an approach which uses the multichannel compressive sensing theory for blind speech separation. The linear programming and gradient-based methods are used to separate the sources. Compared with the conventional blind speech separation, the proposed approach can reduce the requirements of sampling speed and operating rate of the devices. Moreover, our approach has lower computational complexity. The main contribution of this paper lies in proposing a novel procedure to estimate the sources from the measurements without reconstructing the mixed signals. Simulation results demonstrate the proposed algorithm can separate multichannel speech sources successfully.