Audio thumbnailing in video sharing sites

This paper presents a variant of the Smith and Waterman algorithm that operates adaptively on a continuous feature space of MPEG-7 low level spectral descriptors and is capable of detecting repeating patterns (thumbnails) in audio streams that stem from shared Internet videos. The proposed method is not restricted to specific audio types and does not rely on training data. It has been studied in the context of four frequently encountered categories of audio streams, including TV shows, cover versions of music tracks, history documentaries and animal sounds. The results are encouraging and indicate that the presented scheme provides, in the general case, meaningful thumbnails and exhibits acceptable robustness with respect to audio recording quality.