LABROSA ’ S AUDIO CLASSIFICATION SUBMISSIONS

We have submitted a system to MIREX 2008’s audio music classification tasks. It employs the spectral features described in [2] in addition to novel stereo-based features. For the n-way audio classification tasks (artist, classical composer, genre, latin genre, and mood identification) it uses a DAGSVM to perform classification. For the tag classification task, it uses a simple binary SVM with Platt scaling so the SVM outputs for the different tags can be compared. Our system came in first in the audio artist and classical composer identification tasks and tied for first in many of the tag classification and retrieval tasks. A bug in the n-way classification submission prevented the comparison of features, but in the tag classification, the stereo features were shown not to be helpful.