Sparse separation of sources in 3D soundscapes

Abstract

A novel blind source separation algorithm applicable to extracting sources from within 3D soundscapes is presented. The algorithm is based on constructing a binary mask based on directional information. The validity of filtering using binary masked based on the ω-disjoint assumption is examined for several typical scenarios. Results for these test environments show an improvement by an order of magnitude when compared to similar work using speech mixtures. Also presented is the novel application of a dual-tree complex wavelet transform to sparse source separation, providing an alternative transformation to the short-time Fourier transform often used in this area. Results are presented showing compara- ble signal-to-interference performance, and significantly improved signal-to-distortion performance when compared against the short time Fourier transform. Results presented for the separation algorithm include quantitative measures of the separation performance for robust comparison against other separation algorithms. Consideration is given to the related problem of localising sources within 3D sound- scapes. Two novel methods are presented, the first using a peak estimation on a spherical histogram constructed using a geodesic grid, the second by adapting a self learning plastic self-organising map to operate on the surface of a unit sphere. It is concluded that the separation algorithm presented is effective for soundscapes comprising ecological or zoological sources. Specific areas for further work are recog- nised, both in terms of isolated technologies and towards the integration of this work into an instrument for soundscape recognition, evaluation and identification.