Modelling drum patterns with weighted finite-state transducers

In this paper, we present an approach to model bar length drum patterns with weighted finite-state transducers. Motivated by the existing algorithms for speech recognition, we discuss similarities to music, by considering a bar as word and the progression of bars as language. However, in contrast to speech, music has special characteristics, like metrical regularity and multiple notes at one point in time, which have to be taken into account. We use MIDI data to retrieve the drum notes for every bar, which are the input for the training of the WFST models. Once the models are trained, they are used to automatically recognize drum patterns and their time signature in a sequence of unknown drum notes. We present an experiment on symbolic genre classification to demonstrate the principle of operation, training and recognition. It shows that the sequence of drum notes is representative for the four genres: Rumba, Samba, Tango and Waltz leading to a genre recognition rate of 88.9%±2.8%. Applications that could benefit from this approach include drum loop organization, drum note transcription, music similarity and genre detection.