A new auditory-based index to evaluate the blind separation performance of acoustic mixtures

Abstract

A new method to evaluate the performance of convolutive blind signal separation (BSS) algorithms in acoustic mixtures is presented. The method is able to compute the spectral level enhancement, in a frequency-band based fashion, between the estimated and the residual sources, thus combining two previously defined parameters in time and frequency domain: the signal to interference ratio and the spectral preservation index. The new index is able to compute the quality of separation performing the spectral computations over a set of logarithmically spaced frequency bands in a similar way as the human auditory system. Obtained results clearly verify that this methodology is a more realistic approach to evaluate the convolutive BSS results of acoustic mixtures. © Springer-Verlag 2004.