The Development of a Drum Identifier and Transcription Tool

Abstract

Automatic Drum Transcription (ADT) is the process of turning audio recordings of percussion instruments into another representation, typically sheet music or a MIDI file. We propose a novel approach to ADT using a CNN and a generic 2-layer Hierarchical Encoder-Decoder Transformer, which analyzes the drum audio in both the frequency and time axis. While other sub-areas of Automatic Music Transcription (AMT) use Transformers like these, ADT still relies heavily on complex, meticulously tuned CNNs. We evaluated the model with the Expanded Midi Groove Dataset (E-GMD) and IMST dataset and achieved state-of-the-art results with regard to Drum classification, Onset, Offset, and Velocity Estimations.