Perceptual evaluation of musicological cues for automatic song segmentation

The present study evaluated how well boundaries predicted by nine rules, each of them relying on one particular musical cue, can predict perceptual boundaries. The rules were taken from two models: the local boundary detection model (LBDM) by Cambouropoulos (2001) and the generative theory of tonal music (GTTM) by Lerdahl and Jackendoff (1983) in the form quantified by Frankland and Cohen (2004). Furthermore we added the cue timbre change. The predicted boundary profiles from each rule were correlated with perceptual boundary profiles of six songs obtained in a previous study (Bruderer, McKinney, & Kohlrausch, 2009). The individual rule having the highest correlation with the perceptual boundaries was the LBDM-onset. The optimal combination of three rules results in the combination of LBDM-onset, GTTM-rest, and timbre change, yielding a physical correlation of 0.80 to 0.89 between perceptual and model boundary profiles. Analysis of the perceptual cues given for salient boundaries not predicted by the model suggests that incorporating tempo change and harmonic progression could improve the model predictions. The optimal rule combination for segmentation profiles of polyphonic versions of the same songs as obtained in Bruderer, McKinney, and Kohlrausch (2010) was the combination of the LBDM-onset, timbre change, and the start of a rest. (PsycINFO Database Record (c) 2012 APA, all rights reserved)