Joint Beat & Tatum Tracking from Music Signals

This paper presents a method for extracting two key metrical properties, the beat and the tatum, from acoustic signals of popular music. The method is computationally very efficient while performing comparably to earlier methods. High efficiency is achieved through multirate accent analysis, discrete cosine transform periodicity analysis, and phase estimation by adaptive comb filtering. During analysis, the music signals are first represented in terms of accentuation on four frequency subbands, and then the accent signals are transformed into periodicity domain. Beat and tatum periods and phases are estimated in a probabilistic setting, incorporating primitive musicological knowledge of beat–tatum relations, the prior distributions, and the temporal continuities of beats and tatums. In an evaluation with 192 songs, the beat tracking accuracy of the proposed method was found comparable to the state of the art. Complexity evaluation showed that the computational cost is less than 1% of earlier methods. The authors have written a real-time implementation of the method for the S60 smartphone platform.