Unsupervised Feature Selection for Phoneme Sound Classification using Genetic Algorithm

This paper proposes a new method based on Genetic Algorithm for feature selection in phonemes sound classification. Biological studies have shown that human’s ear is sensitive to different resonant frequencies because of ear’s hair cells. Thus, we propose a technique in which genetic algorithm is used to extract audio features similar to human’s ear in order to achieve better classification. In this paper, genetic algorithm is used in order to select appropriate individual’s features in order to classify sound signals accurately. Each individual consists of genes indicating the resonant frequencies inspired from human cochlea hair cells. Then, feature extraction is done by using individual’s information. Moreover, a fitness function by using classification method based on nearest neighbor is used in order to evaluate each individual of population. Furthermore, by using the proposed genetic algorithm, best individual’s features can be found. In order to evaluate this proposed method, a database which consists of 500 samples for each 12 different phoneme classes is created in this paper. The proposed algorithm is compared with an existing typical audio feature selection based on MFCC and the proposed algorithm achieves much better classification accuracy in comparison with MFCC based feature selection method. During generations, the fitness value shows remarkable improvement of sound classification accuracy.