Hybrid Method for Automatic Music Labeling

Abstract

Automatic music labeling on large-scale bases is a premise to provide systems of music recommendation, importante subject in digital world, demanding countless research in music information retrieval field. Although there are large-scale musical bases, such as Million Song Dataset (MSD), that have lowlevel label signal audio, descendant from audio signal, they are weakly labelled, that is, songs labels may be incomplete at a high level regarding emotion, vocals and instrument. This work aims to present the Music Label Miner (MLM), a hybrid method based on grouping, genetic algorithm and statistical correlation, which generates mappings and ossible inferences of high-level labels based on audio signal, through the relationship of a Large-scale base, MSD, with a lower-dimensional Ground Truth reference base. By applying the proposed method, it will be possible to label songs automatically, which contain only low-level labels and, by the models generated from the method, reach high-level labels. The method is composed by: (i) selection and preprocessing of MSD high and low level data, (ii) reference data set called CAL500exp, (iii) MSD data grouping, (iv) CAL500exp vectorization, (v) relationship of vectorized and grouping datasets, (vi) statistical correlation, (vii) mapping, (viii) visualization of selected data characteristics, (ix) generation of models and (x) inference of high and low-level labels.