Multi-scale temporal fusion by boosting for music classification

Short-term and long-term descriptors constitute complementary pieces of information in the analysis of audio signals. However, because they are extracted over different time horizons, it is difficult to exploit them concurrently in a fully effective manner. In this paper we propose a novel temporal fusion method that leverages the effectiveness of a given set of features by efficiently combining multi-scale versions of them. This fusion is achieved using a boosting technique exploiting trees as weak classifiers, which has the advantage of performing an embedded feature selection. We apply our algorithm to two standard classification tasks, namely musical instrument recognition and multi-tag classification. Our experiments indicate that the multi-scale approach is able to select different features at different scales and significantly outperforms the mono-scale systems in terms of classification performance.