This paper presents an algorithm for video summarization, Audio Video Maximal Marginal Relevance (AV-MMR), exploiting both audio and video information. It is an extension of the Video Maximal Marginal Relevance (Video-MMR) algorithm which was only based on visual information. AV-MMR iteratively selects segments which best represent unselected information and are non redundant with previously selected information. As for Video-MMR, AV-MMR is a generic algorithm which is suitable for both single and multiple videos with multiple genres. Several variants of AV-MMR are proposed and the best one is identified by experimentation. Besides, a visual representation of the coherence of audio and video information for a set of audio-visual sequences is also proposed.