AN ADAPTIVE SYSTEM FOR MUSIC CLASSIFICATION AND TAGGING (MIREX 2009 SUBMISSION)
Geoffroy G. Peeters · 2009
This extended abstract concerns one of the two systems submitted by IRCAM for participation in the MIREX 2009 classification and tagging tasks. The system is adaptive and can handle both single-label classification tasks (genre, mood, artist) and multilabel tasks (tagging). Adaptability is attained by means of automatic feature and model selection, which are both embedded in the multiple-instance binary relevance learning of a Support Vector Machine. We propose a criterion function for SVM parameter selection that takes into account unbalanced sets and the effects of overfitting. The same algorithm, without any manual parameter adaptation, was submitted to all classification tasks. However, it was evaluated in two different configurations (also in all classification tasks) related to two different temporal modeling methods: in the first mode (“file”) each track is represented by a single feature vector and in the second (“tw”) texture windows of fixed length are computed, with a later temporal decision fusion.