Feature Extraction from Speech and Phoneme-recognition Using

Speech is a time varying signal. What makes it more interesting is that the information contained in the signal is very difficult to analyze. Traditional methods of speech analysis use Short Time Fourier Transform and Mel Frequency Cepstral Coefficients (MFCC) to extract the feature out of a speech signal, and the model has been successfully implemented in many speech recognition machines. Their performance varies depending upon the context and the results are promising only when the speech signals are of short duration. After coming across the theory of wavelets I learnt that wavelet transform has better Time-Frequency Localization property than STFT [5]. This project report is an excursion into one of the promising application of wavelets, i.e. speech recognition. Here I applied the concept of wavelet transform for feature extraction out of speech signals. To start with I have taken speech signals of small duration, all of my speech signals are phonemes (fundamental unit of speech). To measure the performance of feature extraction I fed the processed speech signal vectors to a neural network for classification. For classification I have used the standard Self Organizing Feature Map (SOFM) [11] which is an unsupervised learning method. The results I found are interesting and require further experimentation which will be continued in future.