A Physiologically Motivated Approach to the Classification of Natural Sounds using High Order Sound Statistics

The ability of humans and animals to classify sounds into behaviorally relevant categories is an ongoing field of study in auditory neuroscience and psychophysics. We employed a physiologically motivated signal processing scheme to extract time-varying statistics from a large database of real sounds. Our hypothesis is that two sound statistics, intensity and contrast, will be sufficient to classify broad categories of sounds using a naïve Bayesian classifier. We seek to evaluate this hypothesis, determine to what extent these two time-varying statistics can be used to classify sounds and to evaluate the limitations of the technique. The sounds themselves were organized into hierarchical categories based on the species or physical phenomenon generating the sound. Results of the naive Bayesian classifier suggest that classification using these statistics is better than chance for 13 different categories and in some cases has accuracy well above 50%. Performance generally increases with increasing duration of the validation sounds. Sounds from similar sources, such as flowing water and rain or two types of birds are often confused for each other. The classifier has high performance when comparing sounds at a much higher level of hierarchical classification, such as animal vocalizations compared to non-animal environmental sounds. Alternately, categories composed of more disparate sound sources, such as new world primates, have comparatively poor performance. This suggests that contrast and intensity statistics provide critical information that can contribute to sound categorization and that the hierarchical approach to classification is appropriate for many, but not all, types of sounds