利用上下文信息的网球赛声学事件检测

Audio Event Detection in Tennis Match Based on Contextual Information

以检测网球比赛中的声学事件为目的,提出了一种基于上下文信息的声学事件检测方法.该方法利用运动员击球时的吼叫声提供的上下文信息检测击球和裁判判球这两类声学事件;采用非监督学习的方式,充分利用了运动员的吼声为击球声学事件建立模型.与目前的方法相比,所提出的方法不需要人工标注的数据,能有效消除声学事件检测中训练和识别失配所带来的问题.实验表明,所提出的方法能进一步提升声学事件检测的准确率.

This paper sets its goal to audio event detection in tennis match and proposes an audio event detec- tion method based on contextual information. This method detects the sounds of ball hit and line judge＇ s shout by making use of the grunts of players when they hit the ball. In the construction of models for audio events, the unsu- pervised learning is employed to use the information of grunts. Compared with current methods, the proposed meth- od does not need any labeled data, and can reduce the mismatch between training and testing. Experimental results further conform that the proposed method can improve the performance of audio event detection substantially.