Automatic Drum Transcription Using Bi-Directional Recurrent Neural Networks

Automatic drum transcription (ADT) systems attempt to generate a symbolic music notation for percussive instruments in audio recordings. Neural networks have already been shown to perform well in fields related to ADT such as source separation and onset detection due to their utilisation of time-series data in classification. We propose the use of neural networks for ADT in order to exploit their ability to capture a complex configuration of features associated with individual or combined drum classes. In this paper we present a bi-directional recurrent neural network for offline detection of percussive onsets from specified drum classes and a recurrent neural network suitable for online operation. In both systems, a separate network is trained to identify onsets for each drum class under observation—that is, kick drum, snare drum, hi-hats, and combinations thereof. We perform four evaluations utilising the IDMT-SMT-Drums and ENST minus one datasets, which cover solo percussion and polyphonic audio respectively. The results demonstrate the effectiveness of the presented methods for solo percussion and a capacity for identifying snare drums, which are historically the most difficult drum class to detect.