Signal segmentation and its application in the feature extraction of speech

Abstract

Speech is considered as a time-varying signal since the parameters of the signal such as the amplitude, frequency and phase varies in time. Segmenting a duration of captured speech into analysis frames of 20 msecs ensures the assumption of stationarity. If a captured speech segment representing a word that may last for 600 msec, then a total of 30 analysis frames are required to the word. Due to the possibility that adjacent frames are identical, then it would be of interest to combine these frames into a single long frame. The interval where adjacent frames have identical parameters is referred as the time-invariant interval (TII). It is of interest to determine these intervals and two methods presented are the instantaneous energy and frequency estimation (IEFE) and localized time correlation (LTC) function. A comparison is made in the accuracy in the TII estimate for a set of speech samples.