Posted on 12 December 2012
Part 9 in a series of videos recorded from ACM MIRUM 2012 in Nara, Japan.
When assessing similarity between two songs, most music recommender systems treat every portion of a song equally. However, humans clearly pay more attention to certain parts of a song than others. Therefore, one might hypothesize that a music recommender which weights the parts of a song that are more prominent or salient more heavily should return more perceptually relevant results than a music recommender which does not incorporate salience at all.
Mark Cartwright tests this hypothesis through experiments where human subjects are asked to listen to synthesized sounds. First, the authors test how well one auditory cue, novelty, affects temporal salience. Second, the authors test if a modified similarity measure, one which incorporates a novelty measure, is more representative of human assessments of audio similarity than a baseline similarity measure which does not incorporate any temporal salience measure.