Compressively sampled speech: How good is the recovery?

Abstract

Modern signal acquisition technologies are made possible by the Nyquist-Shannon sampling theorem (NST). However, this paradigm is extremely wasteful as the signal is compressed before storing it by systematically discarding imperceptible information. Compressive sensing (CS) aims to directly sense the relevant information. Current literature focus either on formulating more computationally-efficient algorithms, or methods which improve the reconstruction quality. In this paper, we quantify the reconstruction quality of compressively sampled speech with a perceptually intuitive metric–the Perceptual Evaluation of Speech Quality (PESQ)–and with the standard average segmental SNR (SNR_seg). The quality of recovery of compressively sampled speech evaluated using PESQ is dependent on the compression ratio, and independent of the number of subbands used to represent the signal in the spectrogram domain.