Audio enhancement and intelligent classification of household sound events using a sparsely deployed array.

A household sound event classification system consisting of an audio localization and enhancement front-end cascaded with an intelligent classification back-end is presented. The front-end is composed of a sparsely deployed microphone array and a preprocessing unit to localize the source and extract the associated signal. In the front-end, a two-stage method and a direct method are compared for localization. The two-stage method introduces a subspace algorithm to estimate the time difference of arrival, followed by a constrained least squares algorithm to determine the source location. The direct localization methods, the delay-and-sum beamformer, the minimum power distortionless response beamformer, and the multiple signal classification algorithm are compared in terms of localization performance for sparse array configuration. A modified particle swarm optimization algorithm enabled an efficient grid-search. A minimum variance distortionless response beamformer in conjunction with a minimum-mean-square-error postfilter is exploited to extract the source signals for sound event classification tasks that follow. The back-end of the system is a sound event classifier that is based on convolutional neural networks (CNNs), and convolutional long short-term memory networks Mel-spectrograms are used as the input features to the CNNs. Simulations and experiments conducted in a live room have demonstrated the strength and weakness of the direct and two-stage methods. Signal quality enhancement using the array-based front-end proves beneficial for improved classification accuracy over a single microphone.

Iterative algorithm for solving acoustic source characterization problems under block sparsity constraints.

In this paper, an iterative Compressive Sensing (CS) algorithm is proposed for acoustical source characterization problems with block sparsity constraints. Source localization and signal separation are accomplished in a unified CS framework. The inverse problem is formulated with the Equivalent Source Method as a linear underdetermined system of equations. As conventional approaches based on convex optimization can be computationally expensive and fail to deal with continuously distributed sources, the proposed approach that is adapted from the Newton's method and is augmented with a special pruning procedure is capable of solving the inverse problem far more efficiently with comparable accuracy. The pruning procedure employs a binary mask that admits sparsity constraints of two-dimensional block sources. The binary mask is heuristic in that it tends to promote nonzero positive source magnitudes. In each iteration, the source amplitude vector is on one hand updated by the Newton's method and on the other hand pruned with the binary mask. With the pruning procedure, the source magnitudes become increasingly sparse and clustered such that the block characteristics are enhanced. In the post-processing phase, particle velocity is calculated on the basis of the equivalent source amplitudes. Numerical and experimental investigations are conducted to validate the proposed technique. The results have demonstrated the efficacy of the proposed Compressive Newton's method in imaging block sources and extracting signal waveforms with little computational cost, as compared to a convex optimization package.