Improving Automatic Drum Transcription Using Large-Scale Audio-To-Midi Aligned Data

One of the major challenges in Automatic Drum Transcription (ADT) research is the lack of large-scale labeled dataset featuring audio with polyphonic mixtures; this limitation around data availability greatly impedes the progress of data-driven approaches in the context of ADT. To tackle this issue, we propose a semi-automatic way of compiling a labeled dataset using the audio-to-MIDI alignment technique. The resulting dataset consists of 1565 polyphonic mixtures of music with audio-aligned MIDI ground truth. To validate the quality and generality of this dataset, an ADT model based on Convolutional Neural Network (CNN) is trained and evaluated on several publicly available datasets. The evaluation results suggest that our proposed model, which is trained solely on the compiled dataset, compares favorably with the state-of-the-art ADT systems. The result also implies the possibility of leveraging audio-to-MIDI alignment in creating datasets for a broader range of audio-related tasks.