The Research Of An Adaptive Speech Endpoint Detection Tecnology

In this paper, we firstly introduce several speech endpoint detection methods which are wildly used currently and discuses their advantages and limitations. Then we summarize some important measures which are widely used in speech endpoint detection technologies such as short time energy, short time zero-crossing-rate, and short time entropy. Moreover, we propose a new audio feature, EZEf, which combines the advantages of time-domain and frequency-domain criteria and is more robust in noisy environment. Based on this new feature, we propose an adaptive endpoint detection method which makes the EZEf adapt the unanticipated changes of the background by analyzing the background noise timely. This endpoint detection method can be used in the cases of long utterance and frequently changed environment. Finally, we design and implement a caption auto-generation system and the performance of the adaptive endpoint detection algorithm described in this paper is proved well in the application.