A comparison of BSS algorithms in harsh environments

This paper compares the performance of several blind source separation (BSS) algorithms in environments of varying reverberation, noise, microphone spacing, and sparsity. Of particular interest are two frequency domain algorithms; one Cascaded ICA with Intervention Alignment (CICAIA), and one algorithm by Pham, Servière, and Boumaraf. The former is found to work exceptionally well in high noise, low microphone spacing environments. The latter proves to work exceptionally well in high SNR and moderate- to widely-spaced arrays. In addition, while the literature on BSS algorithms is extensive, their performance under varying noise conditions has not been widely explored. Also, though usually BSS results for given reverberation times are provided, the sparseness of the room responses (with the same reverberation times) are not. Here we empirically demonstrate that the sparseness of the source-to-sensor impulse responses dramatically effects BSS performance and therefore should always be reported.