Endpoint detection in noisy environment using complexity measure

Abstract

Speech endpoint detection continues to be a challenging problem particularly for speech recognition in noisy environments. The most popular existing detection method is the simple energy detector which performs adequately for clean speech. Problems arise in noisy environments for low energy phonemes at the endpoints. In this paper, we propose a new algorithm based on the theory of fractals and chaos, which is used widely in nonlinear time series analysis techniques. In this proposed method, we applied KC computation complexity into the speech endpoint detection, to achieve excellent overall results. In particular this method is able to reliably detect the onset and offset of speech even for low SNR.