Composer Recognition Based on 2D-Filtered Piano-Rolls

We propose a method for music classification based on the use of convolutional models on symbolic pitch–time representations (i.e. piano-rolls) which we apply to composer recognition. An excerpt of a piece to be classified is first sampled to a 2D pitch–time representation which is then subjected to various transformations, including convolution with predefined filters (Morlet or Gaussian) and classified by means of support vector machines. We combine classifiers based on different pitch representations (MIDI and morphetic pitch) and different filter types and configurations. The method does not require parsing of the music into separate voices, or extraction of any other predefined features prior to processing; instead it is based on the analysis of texture in a 2D pitch–time representation. We show that filtering significantly improves recognition and that the method proves robust to encoding, transposition and amount of information. On discriminating between Haydn and Mozart string quartet movements, our best classifier reaches state-of-the-art performance in leave-one-out cross validation.