This paper describes a method to cluster and synchronize large scale audio-video sequences recorded by multiple users during an event. The proposed method is designed to jointly cluster audio content and synchronize sequences in each cluster to create a multi-view presentation of the event. The method is roughly based on cross-correlation of local audio features. In this paper, three main contributions are presented to obtain a scalable and accurate framework. First, a salient representation of features is used to reduce the computation complexity while maintaining high performance. Second, an intermediate clustering step is introduced to limit the number of comparisons required. Third, a voting approach is proposed to avoid tuning thresholds for cross-correlation. This framework was tested on 164 YouTube concert videos and results demonstrated the efficiency of the method with a correct clustering of 98.8% of the sequences.