A Mid-Level Representation for Capturing Dominant Tempo and Pulse Information in Music Recordings

Automated beat tracking and tempo estimation from music recordings become challenging tasks in the case of nonpercussive music with soft note onsets and time-varying tempo. In this paper, we introduce a novel mid-level representation which captures predominant local pulse information. To this end, we first derive a tempogram by performing a local spectral analysis on a previously extracted, possibly very noisy onset representation. From this, we derive for each time position the predominant tempo as well as a sinusoidal kernel that best explains the local periodic nature of the onset representation. Then, our main idea is to accumulate the local kernels over time yielding a single function that reveals the predominant local pulse (PLP). We show that this function constitutes a robust mid-level representation from which one can derive musically meaningful tempo and beat information for non-percussive music even in the presence of significant tempo fluctuations. Furthermore, our representation allows for incorporating prior knowledge on the expected tempo range to exhibit information on different pulse levels.