Sound classification based on temporal feature integration

The present work contributes to the field of generalized sound classification. We extensively examine the performance of the next three feature sets: a) MPEG-7 Audio Spectrum Projection, b) MFCC (using an alternative method for their extraction) and c) a group derived utilizing critical band based wavelet packets. Subsequently three types of temporal feature integration strategies are applied on the extracted instant values: a) short-term statistics, b) spectral moments and c) two autoregressive functions. During the experimental phase, we organize ten sound classes using professional sound effects collections of high quality. The density of each category is approximated with left-right hidden Markov models. Comparable results with respect to all the feature sets as well as integration methods are provided, which demonstrate the superiority of the short-term statistics method.