A Speech Endpoint Detection Based on Dynamically Updated Threshold of Box-Counting Dimension

Accurate endpoint detection is important for speech procession. The endpoint detection problem is nontrivial for non-stationary backgrounds where noises may be introduced by the speaker, the recording environment and the transmission system. In this paper, an effective endpoint detection algorithm is proposed for improving speech signal processing performance in noisy environment. The proposed speech/pause discrimination method is based on the box-counting dimension. At the same time, a dynamically updated threshold and adaptive window are used to improve the performance of the algorithm. It is characterized by higher accuracy or flexibility, faster processing speed and less computation. Through a large number of samples for laboratory experiments, the results show that the improvements in detection accuracy over representative endpoint detection algorithms for robust speech processing.