On objective feature selection for affective sounds discrimination

We present an objective acoustic feature selection for automatic affective sounds detection based on stochastic evolutionary optimization algorithms. Particle Swarm Optimization (PSO) as well as Genetic Algorithms (GA) are exploit to select the most appropriate audio features from a large set of available features. We performed experiments on a dataset containing about two hours of affective sounds - cry, laughter and applause, and supplemented with several hours of recordings of other sounds (speech, music and various types of noise). Applying the feature selection methods, the classification performance is increased about 4-9 % with final accuracy 92-98 % while feature space dimension is reduced about 50-90 %.