Additional Evidence That Common Low-level Features of Individual Audio Frames Are Not Representative of Music Genre

The Bag-of-Frames (BoF) approach has been widely used in music genre classification. In this approach, music genres are represented by statistical models of low-level features computed on short frames (e.g. in the tenth of ms) of audio signal. In the design of such models, a common procedure in BoF approaches is to represent each music genre by sets of instances (i.e. frame-based feature vectors) inferred from training data. The common underlying assumption is that the majority of such instances do capture somehow the (musical) specificities of each genre, and that obtaining good classification performance is a matter of size of the training dataset, and fine-tuning feature extraction and learning algorithm parameters. We report on extensive tests on two music databases that contradict this assumption. We show that there is little or no benefit in seeking a thorough representation of the feature vectors for each class. In particular, we show that genre classification performances are similar when representing music pieces from a number of different genres with the same set of symbols derived from a single genre or from all the genres. We conclude that our experiments provide additional evidence to the hypothesis that common low-level features of isolated audio frames are not representative of music genres.