Tempo Detection of Urban Music Using Tatum Grid Non Negative Matrix Factorization

High tempo detection accuracies have been reported for the analysis of percussive, constant-tempo, Western music audio signals. As a consequence, active research in the tempo detection domain has been shifted to yet open tasks like tempo analysis of non-percussive, expressive, or non-western music. Also, tempo detection is included in a large range of music-related software. In DJ software, features like beat-synching or tempo-synchronized sound effects are widely accepted in the DJ community, and their users rely on correct tempo hypothesis as their basis. In this paper, we are evaluating both academic and commercial tempo detection systems on a typical dataset of an urban club music DJ. Based on this evaluation, we identify octave errors as a problem that has not yet been solved. Further, an approach based on non-negative matrix factorization is presented. In its current state it can compete with the state of the art. It further provides a foundation to tackle the octave error issue in future research.