Research on Endpoint Detection of Sound Events in Elevator Operation Environment

Research on Endpoint Detection of Sound Events in Elevator Operation Environment

연구 분야: Safety

학회: AIPR '21: Proceedings of the 2021 4th International Conference on Artificial Intelligence and Pattern Recognition

The accuracy and efficiency of sound event recognition are affected by the accuracy of endpoint detection. In a complex environment, it is difficult to detect the endpoint of a sound event due to the influence of background noise. Aiming at the sound events in the elevator operating environment, this paper analyzes and studies multiple characteristic parameters based on short-term energy, short-term average zero-crossing rate, and cepstral distance, and proposes a new characteristic parameter, namely multiple cepstral distance, which is adopted after adding a smoothing mechanism. The new decision mechanism performs endpoint detection. This paper collects environmental sounds of elevators in schools, communities, and shopping malls, and conducts endpoint detection comparative experiments on four sound events: speech, explosion, glass breaking, and alarm sound under different signal-to-noise ratios. The experimental results show that the method can be well adapted to the endpoint detection of four sound events in different environments. Even in the -5dB environment, the average misdetection rate of the four sound events can still be lower than 10%, and the method is robust Better, it has a wide range of application prospects in an elevator safety inspection.