Structured Dictionary Learning for Compressive Speech Sensing

Abstract

Sparse dictionary learning aims at training appropriate redundant dictionaries for specific tasks of signal processing, such as signal estimation, compression and classification. Most of the existing dictionary learning algorithms for compressive speech sensing only exploit speech samples to construct the dictionary. In this paper, we propose to leverage both the speech signal and its linear prediction coefficients jointly to learn a structured and sparse dictionary. The proposed dictionary is designed based on a new optimization strategy using both $l_{0}$ and $l_{2}$ norms to enforce sparsity and structure, respectively. The resulting optimization problem can be solved by a fast iterative algorithm in two stages. Experimental results indicate that our proposed algorithm converges faster than the reference methods while yielding a better objective evaluation performance in terms of segmental signal-to-noise ratio, perceptual evaluation of speech quality and short-time objective intelligibility of the reconstructed speech.