AcousticBrainz Genre Task: Content-based music genre recognition from multiple sources
Over the past months, we’ve been preparing a genre recognition task based on the vast amounts of music data we gathered in AcousticBrainz database. It is now a part of MediaEval 2017, a benchmarking initiative that organizes an annual cycle of scientific evaluation tasks in the area of multimedia access and retrieval.
The task is about music genre recognition: we want to build systems that are able to predict genre and subgenre of unknown music recordings (songs) given automatically computed music audio features of those recordings.
It is a popular problem in Music Information Retrieval, however the task that we propose is somewhat different, more detailed and more challenging:
There are different genre taxonomies and people may not always agree on the meaning of genres. Genres labels are probably subjective categories. We want to explore how the same music can be annotated differently by different communities following different genre taxonomies, and how this should be addressed by genre recognition systems. We provide four genre sources that come from different music databases. Their taxonomies vary in specificity, breadth and meaning of genre labels. These sources include explicit annotations done by music experts and annotations inferred from folksonomies.
Typically research is done on a small number of broad genre categories. In contrast, we propose to consider more specific genres and subgenres and our data contains hundreds of subgenres.
Genre recognition is often treated as a single category classification problem, which is not necessarily the way it should be. Our genre data is intrinsically multi-label and so we propose to treat genre recognition as a multi-label classification problem.
Typically research is done on small music collections. Instead, we provide a very large dataset counting two million recordings annotated by genres and subgenre. The downside is that we are not able to provide audio, but only precomputed bags of features.
Finally, we provide information about the hierarchy of genres and subgenres within each genre annotation source. Systems can take advantage of this knowledge.