ABSTRACT
The sound of rain or of a large crowd are examples of sound textures. A restricted definition of sound texture is proposed for machine processing. Sound textures are treated as two-level phenomena: simple sound elements called atoms form the low level, and the distribution and arrangement of atoms form the high level. A cluster-based probability model is used to characterize the high level of sound textures. The model is then used to resynthesize textures that are perceptually similar to originals (training data). Finally, applications of the model for classification of sound textures are suggested.
