Adaptive compressed sensing of speech signal based on data-driven dictionary

Abstract

Compressed Sensing (CS) is an emerging signal acquisition theory that provides a universal approach for characterizing signals which are sparse or compressible on some basis at sub-Nyquist sampling rate. This paper focuses on the realization of CS on natural speech signals. We construct an over-complete data-driven dictionary as the sparse basis specialized for speech signals. Based on this, CS sampling and reconstruction of speech signal are realized. Furthermore, we propose to choose the sensing matrix adaptively, according to the energy distribution of original speech signal. Experimental results show significant improvement of speech reconstruction quality by using such adaptive approach against using traditional random sensing matrix.