Recurrent Neural Network-Based Dictionary Learning for Compressive Speech Sensing

We propose a novel dictionary learning technique for compressive sensing of speech signals based on the recurrent neural network. First, we exploit the recurrent neural network to solve an l0-norm optimization problem based on a sequential linear prediction model for estimating the linear prediction coefficients for voiced and unvoiced speech, respectively. Then, the extracted linear prediction coefficient vectors are clustered through an improved Linde-Buzo-Gray algorithm to generate codebooks for voiced and unvoiced speech, respectively. A dictionary is then constructed for each type of speech by concatenating a union of structured matrices derived from the column vectors in the corresponding codebook. Next, a decision module is designed to determine the appropriate dictionary for the recovery algorithm in the compressive sensing system. Finally, based on the sequential linear prediction model and the proposed dictionary, a sequential recovery algorithm is proposed to further improve the quality of the reconstructed speech. Experimental results show that when compared to the selected state-of-the-art approaches, our proposed method can achieve superior performance in terms of several objective measures including segmental signal-to-noise ratio, perceptual evaluation of speech quality and short-time objective intelligibility under both noise-free and noise-aware conditions.