Noisy Speech Endpoint Detection Using Robust Feature

In this paper a new robust feature for speech endpoint detection is proposed. It combines the properties of the Modified Group Delay Spectrum (MGDS) and the Mean Delta (MD) approach in order to obtain the more robust endpoint detection. This feature is named as Group Delay Mean Delta (GDMD) feature. The effectiveness of proposed feature and other three features for trajectory- based endpoint detection is experimentally evaluated in the fixed-text Dynamic Time Warping (DTW) - based speaker verification task with short phrases of telephone speech. The analysed features are - Modified Teager Energy (MTE), Energy-Entropy (EE) feature and MD feature. The results of the experiments have shown that the GDMD feature demonstrates the best performance in endpoint detection tests in terms of verification rate.