GIF-SP: GA-based Informative Feature for Noisy Speech Recognition

This paper proposes a novel discriminative feature extraction method. The method consists of two stages; in the first stage, a classifier is built for each class, which categorizes an input vector into a certain class or not. From all the parameters of the classifiers, a first transformation can be formed. In the second stage, another transformation that generates a feature vector is subsequently obtained to reduce the dimension and enhance recognition ability. These transformations are computed applying genetic algorithm. In order to evaluate the performance of the proposed feature, speech recognition experiments were conducted. Results in clean training condition shows that GIF greatly improves recognition accuracy compared to conventional MFCC in noisy environments. Multi-condition results also clarifies that out proposed scheme is robust against differences of conditions.