Discriminative Feature Selection for Applause Sounds Detection

The specific sounds such as applause, laughter, explosions, etc. are very helpful to understand high level semantic of audio/video content. The paper focuses on feature selection by evolutional programming for an automatic detection of applause in audio stream. A set of the most discriminative features is selected by Genetic Algorithm and Simulated Annealing. The experiments are run on more than 9 hours of audio selected from various audio and video content. The results show that the applause sound recognition improves if only a few coefficients are selected from MFCC static and dynamic features. Further, the delta-delta coefficients (the 2nd time derivates of MFCCs) highly outperform the delta coefficients.