Feature dimensionality reduction through Genetic Algorithms for faster speaker recognition

Mel-Frequency Cepstral Coefficients and their derivatives are commonly used as acoustic features for speaker recognition. Reducing the number of features leads to more robust estimates of model parameters, and speeds up the classification task, which is crucial for real-time speaker recognition applications running on low-resource devices. In this paper, a feature selection procedure based on Genetic Algorithms (GA) is presented and compared to two well-known dimensionality reduction techniques, namely PCA and LDA. Evaluation is carried out for two speech databases, containing laboratory read speech and telephone spontaneous speech, applying a standard speaker recognition system. Results suggest that dynamic features are less discriminant than static ones, since the low-size optimal subsets found by the GA did not include dynamic features. GA-based feature selection outperformed PCA and LDA when dealing with clean speech, whereas PCA and LDA outperformed GA-based feature selection for telephone speech, probably due to some kind of noise compensation implicit in linear transforms, which cannot be accomplished just by selecting a subset of features.