Speech recognition by clustering wavelet and PLP coefficients

This thesis explores the use of K-means clustering, wavelets and perceptual linear predictive (PLP) analysis for speech recognition. First, we would like to compare the performance of this method to those of previous speech recognition techniques. Next, we want to test if both wavelet and PLP coefficients are important to the analysis. Finally, we want to try out various means of improving the clustering method. The task of phonetic classification is used as the basis of comparison. Three different sets of phonemes are chosen from the TIMIT database: 16 vowels, 24 consonants and 39 phonetic classes. Coefficients of the Haar wavelet transform and the 5th order PLP analysis are combined to form a 42-dimensional vector for each phoneme. Clusters of phoneme vectors obtained by K-means clustering are then used to classify test vectors. Classification experiments using the NIST train and test sets show that independent clustering of phonemes with proportional phoneme emphasis is the best clustering strategy. It yields an accuracy of 55.4% for vowel classification, 54.6% for consonant classification and 50.9% for phoneme classification. Tests with waveletonly vectors and PLP-only vectors show that both the wavelet transform and PLP analysis are significantly important to the phonetic classifier. Results also show that wavelet coefficients are useful for detecting sound transitions, which are abundant in consonants. Thesis Supervisor: Kenneth Yip Title: Visiting Assistant Professor