Separation of acoustical source power spectral densities with Bayesian sparsity enforcing

Abstract

Active research is ongoing to improve the design of different patterns of aircrafts including innovative devices of noise reduction often assessed during experiments conducted with scaled models in wind tunnels or in situ with real aircrafts. Source localization methods play a fundamental role to identify the source locations which are at the origin of the annoyance. Another topic for these experiments is the establishment of theoretical models, requiring a fine picture of the Power Spectral Densities (PSDs) of the main sound sources. The topic is to extract the PSDs of the primary source signals from, the PSDs of the measured mixtures. Blind signal separation techniques seem to be suited for this problem. Among the numerous existing methods, the Bayesian separation approach has the advantage of incorporating relevant information about the PSDs of the sources, and the mixing systems to help the separation process. This approach is enforced to the separation of the PSDs of primary source signals recorded by an array of microphones during tests performed in an anechoic chamber with tonal, narrow-band and broadband acoustic sources. The Signal-to-Distortion Ratio (SDR) allows to show that the separation results are better when sparsity priors are used to describe the source PSDs rather than Gaussian ones for all the scenarios of mixtures considered in the article. We demonstrate that the SRD decreases in a similar manner as the measure of the sparseness of the PSDs of the acoustic sources.