Compressive Blind Mixing Matrix Estimation of Audio Signals

Abstract

Compressive sensing (CS) shows that, when a signal is sparse or compressible with respect to some basis, a small number of compressive measurements of the original signal can be sufficient for exact (or approximate) recovery. Distributed CS (DCS) takes advantage of both intra- and intersignal correlation structures to reduce the number of measurements required for multisignal recovery. In most cases of audio signal processing, only mixtures of the original sources are available for observation under the DCS framework, without prior information on both the source signals and the mixing process. To recover the original sources, estimating the mixing process is a key step. The underlying method for mixing matrix estimation reconstructs the mixtures by a DCS approach first and then estimates the mixing matrix from the recovered mixtures. The reconstruction step takes considerable time and also introduces errors into the estimation step. The novelty of this paper lies in verifying the independence and non-Gaussian property for the compressive measurements of audio signals, based on which it proposes a novel method that estimates the mixing matrix directly from the compressive observations without reconstructing the mixtures. Numerical simulations show that the proposed method outperforms the underlying method with better estimation speed and accuracy in both noisy and noiseless cases.