Audio Mood Classification Using Ensemble Classifier with Music Tag Based Indexing

This paper presents a system for audio classification using multiple binary ensemble classifiers with music tag based indexing in a one-versus-all classification scenario. The proposed system has won the audio classification task on mood dataset in MIREX 2010 and is implemented as follows. First, in the training phase, the frame-based 70-dimensional feature vectors are extracted from a training audio track by MIRToolbox 1.3. Next, the Posterior Weighted Bernoulli Mixture Model (PWBMM) is applied to index the frame-decomposed feature vectors of the training track into a fixed-dimensional semantic vector based on the pre-defined music tags, called music tag based indexing. Then for each class, the semantic vectors of associated training tracks are used to train a binary ensemble classifier consisting of SVM and AdaBoost. In the classification phase, a testing audio track is first indexed into a semantic vector. Then, for each class, the associated ensemble classifier gives its final score for the testing track via the calibrated probability ensemble between the two subclassifiers, i.e., SVM and AdaBoost. The class with the highest final score is assigned as the final output. Our system was ranked first out of 36 submissions in the MIREX 2010 audio mood classification task.