End points detection for noisy speech using a wavelet based algorithm

Abstract

This paper represents a way for the detection of start-end boundaries of a speech segment in the presence of noise using the wavelet transform. The technique is based on generating a certain mathematical function derived from the wavelet parameters that can keep track of the energy changes along the speech duration. The problem in end points detection always appears in words that begin or end in low-energy phonemes. This problem can be entirely eliminated by this technique. Examples are given to show how the algorithm performs under different signal-to-noise ratios.