A Statistical Approach to the Design of an Adaptive Self-Normalizing Silence Detector

Abstract

Silence detection is an important aspect of automatic speech recognition. In this paper a silence detector is described which classifies each frame of an utterance as being silent or nonsilent. No attempt is made to determine the endpoints of words. The silence detector is based upon a statistical test and has the properties that it is adaptive, self-normalizing, speaker independent, and script independent. No training sets are required, although it is assumed that the first part of a signal is silent. An experiment is described in which the silence detector was successfully applied to more than 2 h of isolated-word speech using a frame length of 1 cs. © 1983 IEEE