An Algorithm for Determining the Endpoints of Isolated Utterances

The problem of locating the beginning and end of a speech utterance in an acoustic background of silence is important in many areas of speech processing. In particular, the problem of word recognition is inherently based on the assumption that one can locate the region of the speech utterance to be recognized. A further advantage of a good endpoint-locating algorithm is that proper location of regions of speech can substantially reduce the amount of processing required for the intended application. The task of separating speech from background silence is not a trivial one except in the case of acoustic environments with extremely high signal-to-noise ratio, e.g., an anechoic chamber or a soundproof room in which high-quality recordings are made. For such high signalto-noise ratio environments, the energy of the lowest-level speech sounds (e.g., weak fricatives, low-level voiced portions, etc.) exceeds the background noise energy and a simple energy measure suffices.1 However, such ideal recording conditions are not practical for realworld applications of speech-processing systems. Thus, simple energy 297