Using vector of fractal dimensions for feature reduction and phoneme recognition and classification

Abstract

Difference between Hausdorff fractal dimensions of phonemes gave us a motivation to use this feature as input of a statistical Bayesian classification system and a nearest neighborhood (NN) classifier for speech waveform recognition. We divide phoneme waveforms to adjacent segments and calculate Hausdorff fractal dimension of each segment and using them as the input of a Bayesian/Nearest Neighborhood classifier. The power point of algorithm is in consideration of order of samples information in contrast of other non-supervised feature extraction algorithms.