STAR Drums: A Dataset for Automatic Drum Transcription

Abstract

Current state-of-the-art Automatic Drum Transcription (ADT) algorithms make use of neural networks. To train such models, large amounts of annotated data are needed. We introduce the Separate-Tracks-Annotate-Resynthesize Drums (STAR Drums) dataset, derived from full audio recordings that include mixtures of drum instruments, melodic instruments, and vocals. First, we separate the music recordings into a drum stem and a non-drum stem by applying a Music Source Separation (MSS) algorithm, and then automatically annotate the drum stem with an ADT algorithm. The annotations are utilized for the re-synthesis of the drum stem using samplebased virtual drum instruments. Finally, we mix the re-synthesized drum stem with the original non-drum stem to obtain the final mix. In summary, STAR Drums comprises annotated synthesized drum sounds mixed with real recordings of melodic instruments and vocals, offering several benefits: high temporal accuracy of annotations; training data that includes recordings of instruments played by musicians, rather than solely relying on MIDI-rendered audio; a large number of supported drum classes; the possibility to customize the final mix by, for instance, applying additional processing to the drum stem, as both drum and non-drum stems are provided; and suitable licenses of audio files for making the dataset fully available to the research community. We demonstrate that, in the context of ADT, training with STAR Drums achieves superior performance compared to training with datasets solely relying on MIDI-rendered data and that the synthesized nature of the drum stem does not diminish performance.

If you make use of the dataset, please cite the accompanying paper which provides a detailed description: Weber, P., Uhle, C., Müller, M. and Lang, M. (2025) ‘STAR Drums: A Dataset for Automatic Drum Transcription’, Transactions of the International Society for Music Information Retrieval, 8(1), p. 248–264.

Download

For easier file handling, we divided the zipped dataset into 30 GB chunks. To rejoin the parts, use:

cat STAR_Drums_full.zip.part-* > STAR_Drums_full.zip

Then, you can verify the integrity of the created zip file. Save STAR_Drums_full.zip.sha256 in the same directory where STAR_Drums_full.zip is located and run:

sha256sum -c STAR_Drums_full.zip.sha256

This will take a few minutes and should then return ./STAR_Drums_full.zip: OK.

Structure

Annotations for 18 drum classes, audio stems, and MIDI files used to render the re-synthesized drum stems are located in the ./data directory, organized into sub-folders for the training, validation, and test splits.

Each annotation consist of a time stamp in seconds, followed by the drum class, followed by the MIDI velocity value

Scripts: In ./data/scripts, we provide

code to create mix files for the MUSDB-18 items whose licenses do not permit direct distribution.

code as a starting point for creating data-augmented versions of the mixes by pre-processing the re-synthesized drum stems or altering mixing gains.

the midi_mappings which were used to create MIDI files from the annotations using the virtual instruments referred to in the paper