A New Approach to Sound Source Segregation

We rely critically on our ability to 'hear out' (segregate) individual sound sources in a mixture. Yet, despite its importance, little is known regarding this -ability. Perturbation analysis is a psychophysical method that has been successfully applied to related problems in vision [Murray, R.F. 2011. J. of Vision 11, 1-25]. Here the approach is adapted to audition. The application proceeds in three stages: First, simple speech and environmental sounds are synthesized according to a generative model of the sound--producing source. Second, listener decision strategy in segregating target from non--target (noise) sources is determined from decision weights (regression coefficients) relating listener judgments regarding the target to lawful perturbations in acoustic parameters, as dictated by the generative model. Third, factors limiting segregation are identified by comparing the obtained weights and residuals to those of a maximum-likelihood (ML) observer that optimizes segregation based on the equations of motion of the generating source. Here, the approach is applied to test between the two major models of sound source segregation; target enhancement versus noise cancellation. The results indicate a tendency of noise segregation to preempt target enhancement when the noise source is unchanging. However, the results also show individual differences in segregation strategy that are not evident in the measures of performance accuracy alone.