Skip to Main Content
This paper addresses the problem of automatically extracting perceptive information from acoustic signals, in a supervised classification context. Global labels, i.e., atomic information describing a music title in its entirety, such as its genre, mood, main instruments, or type of vocals, are entered by humans. Classifiers are trained to map audio features to these labels. However, the performances of these classifiers on individual labels are rarely satisfactory. In the case we have to predict several labels simultaneously, we introduce a correction scheme to improve these performances. In this scheme-an instance of the classifier fusion paradigm-an extra layer of classifiers is built to exploit redundancies between labels and correct some of the errors coming from the individual acoustic classifiers. We describe a series of experiments aiming at validating this approach on a large-scale database of music and metadata (about 30 000 titles and 600 labels per title). The experiments show that the approach brings statistically significant improvements.