Automatic Music Genre Classification in Small and Ethnic Datasets

Automatic music genre classification commonly relies on a large amount of well-recorded data for model fitting. These conditions are frequently not met in ethnic music collections due to low media availability and ill recording environments. In this paper, we propose an automatic genre classification technique especially designed for small, noisy datasets. The proposed technique uses handcrafted features and a votebased aggregation process. Its performance was evaluated over a Brazilian ethnic music dataset, showing that using the proposed technique produces higher F1 measures than using traditional data augmentation methods and state-of-the-art, Deep Learning-based methods. Therefore, our method can be used in automatic classification processes for small datasets, which can be helpful in the organization of ethnic music collections.