Bio-inspired optimization of acoustic features for generic sound recognition

We propose the generic sound recognition system that exploits evolutional algorithms for a selection of discriminative acoustic features. Namely, we applied Particle Swarm Optimization and Genetic Algorithms to select the most significant acoustic features from a large collection of audio features. The system, which is based on k-Nearest Neighbors algorithm, classifies sounds into the following six classes - speech, music, noise, applause, laughing and crying. The experimental results show that both algorithms give solutions of almost equal quality. Compared to the case when all audio features are used, the proposed optimization process gains improvement in classification accuracy from 72.64 % to 82.48 % and in addition, it makes a reduction of feature space dimension down to 62.77 % of original size.