Endpoint detection based on mel-scale features and phoneme segmentation

The conventional methods for speech endpoint detection based on some simple features such as energy and zero-crossing cannot obtain precise results. The proposed method in this paper is based on combination of Mel-scale features with phoneme segmentation. The experiments show that the high accurate detection can be obtained. In addition, the detector cannot only segment vowel and consonant, but also consonants itself. It can help further to reduce recognition error in speech recognition.