Wavelet method of speech segmentation

In this paper a new method of speech segmentation is suggested. It is based on power fluctuations of the wavelet spectrum for a speech signal. In most approaches to speech recognition, the speech signals are segmented using constant-time segmentation. Constant segmentation needs to use windows to decrease the boundary distortions. A more natural approach is to segment the speech signals on the basis of time-frequency analysis. Boundaries are assigned in places where some energy of a frequency band rapidly changes. Most methods of non-constant segmentation need training for particular data or are realized as a part of modelling. In this paper we apply the discrete wavelet transform (DWT) to analyse speech signals, the resulting power spectrum and its derivatives. This information allows us to locate the boundaries of phonemes. It is the first stage of speech recognition process. Additionally we present an evaluation by comparing our method with hand segmentation. The segmentation method proves effective for finding most phoneme boundaries. Results are more useful for speech recognition than constant segmentation.