Acoustics 08 Paris 3141 The texture of sound sources is a robust and characteristic perceptual property that listeners use for source recognition. This work introduces a method to determine the presence of sound textures associated with noise, pulses and tonal contributions and applies this to sound source similarity prediction. Local textures are estimated by applying a Kolmogorov-Smirnov test to compare the histogram of locally estimated texture information with the histogram for broadband

more »

... se. This method can determine the texture of TF-regions as small as 50 ms by 1 octave. With this measure we analyze cochleograms (i.e. spectrograms based on a cochlea-model) in terms of the presence of three basic textures: broadband noise, tones, and pulses. The relative contributions of the three classes are used as a distance measure between sounds. These distances are compared with the differences between sounds that listeners reported in an experiment by Brian Gygi. Our results lead to an organization of natural sounds similar to the perceptual organization that resulted from Gygi's sound similarity study. We conclude that human sound classification first and foremost is aimed at separating input sounds into broadband noise, pulses and tonal components. Results This section addresses three ways to compare the calculated and the experimental similarity measures. The first measure addresses a direct comparison of the two sets. The second measure relies on MDS to form a more efficient Acoustics 08 Paris