An event sequence based method for audio scene analysis

Audio semantic analysis is an important issue for multimedia applications. In this paper, we propose a neural network based approach to analyze the high-level semantic content of audio event sequences for the action movies. According to the time interval between adjacent basic audio events, we first divide the given event sequence into some scene segments, and then discover the high-level semantic content of the audio context. By using the neural network based approach, the prior knowledge and the machine learning are effectively combined in the semantic inference. Specifically, the model parameters are learned by the statistical learning, and then are modified manually based on the prior knowledge. Moreover, we select some audio streams from the action movies to evaluate the performance of the proposed approach. The experiment results demonstrate the approach can work well.