Options
2010
Conference Paper
Titel
Improving Markov model-based music piece structure labelling with acoustic information
Abstract
This paper proposes using acoustic information in the labelling of music piece structure descriptions. Here, music piece structure means the sectional form of the piece: temporal segmentation and grouping to parts such as chorus or verse. The structure analysis methods rarely provide the parts with musically meaningful names. The proposed method labels the parts in a description. The baseline method models the sequential dependencies between musical parts with N-grams and uses themfor the labelling. The acoustic model proposed in this paper is based on the assumption that the parts with the same label even in different pieces share some acoustic properties compared to other parts in the same pieces. The proposed method uses mean and standard deviation of relative loudness in a part as the feature which is then modelled with a single multivariate Gaussian distribution. The method is evaluated on three data sets of popular music pieces, and in all of them the inclusion of the acoustic model improves the labelling accuracy over the baseline method.