Wavelet Transform Speech Recognition Using Vector Quantization, Dynamic Time Warping And Artificial

In this paper we investigate the performance of the Discrete Wavelet Transform (DWT) with Dynamic Time Warping, Vector Quantization and Artificial Neural Networks for speaker-dependent, isolated word recognition. Wavelet Transforms have demonstrated good time-frequency localization properties and are appropriate tools for the analysis of non-stationary signals like speech. Moreover, unlike LPC, they do not assume any model for the input speech signal. The DWT may also be implemented as a fast, pyramidal algorithm. Our experiments and simulation results indicate that the DWT is a potential contender to be a feature extraction tool for speech recognition. While vector quantization and dynamic time-warping with wavelets have yielded results comparable to those obtained with LPC, preliminary results suggest that the neural networks are sensitive to time alignment and frame synchronism.