Drum Transcription via Joint Beat and Drum Modeling Using Convolutional Recurrent Neural Networks

Existing systems for automatic transcription of drum tracks from polyphonic music focus on detecting drum instrument onsets but lack consideration of additional meta information like bar boundaries, tempo, and meter. We address this limitation by proposing a system which has the capability to detect drum instrument onsets along with the corresponding beats and downbeats. In this design, the system has the means to utilize information on the rhythmical structure of a song which is closely related to the desired drum transcript. To this end, we introduce and compare different architectures for this task, i.e., recurrent, convolutional, and recurrent-convolutional neural networks. We evaluate our systems on two well-known data sets and an additional new data set containing both drum and beat annotations. We show that convolutional and recurrentconvolutional neural networks perform better than state-ofthe-art methods and that learning beats jointly with drums can be beneficial for the task of drum detection.