Unsupervised Detection of Structural Changes in Electronic Dance Music

Music structure segmentation is a relevant topic in music information retrieval as it is both useful to the analysis of musical structure and a means to improve performance in other tasks [1], such as audio editing in recording workflows [2]. Recent years have seen Electronic Dance Music (EDM) gaining access to mainstream charts. However, it is still a neglected genre in music research with just a few articles published in journals and conferences. One exception is Butler [3], who claims that "EDM presents a distinctive overall formal structure in addition to its characteristic instrumentation". We present a method for the location of structural changes in EDM. Our approach starts by detecting the first downbeat of a track. Tempo is then computed using a standard tempo detection algorithm [7]. We then calculate the magnitude spectrum for each frame using a Fast Fourier Transform (FFT). The frames are beat-aligned with 87.5% overlap. Contrarily to most common approaches, we perform a cepstrum analysis in order to find periodic sequences in the signal. Based on the approach by Foote [5], we then compute the cosine distance between each possible pair of frames in the cepstrum data to get a self-similarity matrix. Convoluting along the main diagonal of the similarity matrix using a Gaussian checkerboard kernel yields a unidimensional linearly normalized novelty curve that indicates the temporal locations of significant textural changes. We use a kernel size of approximately 30s. Justifications for this value can be found in [6]. Positions of prominent peaks of the novelty curve are selected as candidates for segment boundaries. Finally we propose a set of heuristic rules to align the obtained novelty peaks with the beats (see Figure 1). In line with Butler [4], who explains how EDM structure relies on sequences of 8 or 16 bars of 4 beats, these heuristics apply an asymmetric weight to these bars. Table 1 shows the results obtained for different datasets using the same parameter settings. It performs well on an in-house EDM dataset. Although this method was created specifically for EDM, results on the RWC Pop dataset [8] are in line with the best performing algorithms submitted to MIREX 2012. This means that structural changes in pop music might have the same periodicity as in EDM. However, this method does not reach high performance on the Eurovision dataset. An explanation for this might be that pop music is mixed with traditional music from European countries in this song contest. Figure 1: Timeline is shown in beats (0 is the first beat detected). Heuristic rules dictate an asymmetric weight towards the 8 and 16 bars.