Abstract: A new method to classify the violent videos by the bag of audio words was introduced. The MPEG7 audio descriptors are firstly extracted, including the low level features such as AudioSpectrumCentroid and AudioSpectrumSpread etc. After that, the audio words are built through the MPEG7 high level descriptor, the AudioSighnature, which is considered as the fingerprint of the audio stream. The support vector machine is used to classify the feature vectors into two genres, which are the violent and nonviolent. There are three experiments in this paper: the research on the different types of the audio words, the different size of words and the classification of the shots detected from the visual features. It is demonstrated from the experiment result that the proposed method achieves good recall accuracy.