Repeating segment detection in songs using audio fingerprint matching

We propose an efficient repeating segment detection approach that doesn't require computation of the distance matrix for the whole song. The proposed framework first extracts audio fingerprints for the whole song. Then,for each time step in the song we perform a query to match a sequence of M fingerprint codewords against the fingerprints of the rest of the song. In order to find a match for the first fingerprint query, a search tree data structure is built with the fingerprints of the rest of the song. For subsequent fingerprint queries for the rest of the song, the matching process dynamically updates the search tree data structure to exclude the M fingerprint codewords corresponding to each time step. For each matching segment, we record the time offset from the query segment. Following the matching process for the whole song, we compute the histogram of the number of matching segments for each offset. The peaks in this histogram correspond to offsets at which matches were found more often than others and can be used to pick out a set of repeating segments.