D-Composer: Language Modeling for Simultaneous Drum Transcription and Sound Event Decomposition

One-shot Decompositions

Given a drum recording, D-Composer predicts a transcript and isolated one-shots corresponding to every note in the transcript. The baseline Inverse Drum Machine (IDM) system predicts a transcript and a single isolated one-shot per drum class. Taken together, these predictions can be thought of as "decompositions" of drum recordings into their constituent sound events.

For both D-Composer and IDM, we arrange the predicted one-shots according to the predicted transcripts in a playable piano roll format. We provide examples across five datasets, spanning both acoustic and electronic drum timbres. Click the arrows to toggle between examples within each dataset, click Play to listen to the "round-trip" reconstruction obtained by synthesizing the decomposition as audio, and click on notes to listen to individual one-shot predictions.

All one-shots in the piano rolls below are synthesized. D-Composer synthesizes unique one-shot audio for each note event, while IDM synthesizes one-shot audio for each drum class and replicates this audio (modulo volume) across all note events of the same class.

Round-Trip Reconstructions

Additional examples of "round-trip" reconstructions synthesized from D-Composer and IDM decompositions. For each reconstruction, we report the multi-scale spectral loss (MSL) vs. ground truth.

One-Shot Extraction Examples

We provide one-shot extraction examples for the proposed D-Composer and baseline IDM and DOSE systems. Each input mixture is synthesized from known ground-truth one-shots of each drum class. Where the proposed or baseline methods do not predict a one-shot of a given class, we display "missing". For each predicted one-shot, we report the multi-scale spectral loss (MSL) vs. ground truth.

Codec Reconstruction Examples

We provide example one-shot reconstructions from the proposed DOC, the baseline codecs DAC and CoDiCodec, and the continuous MelodyFlow VAE on which DOC operates. For each reconstructed one-shot, we report the multi-scale spectral loss (MSL) vs. ground truth.

Notes / Corrections

MusDB Dataset

In Section 4.3 of the paper, we describe our evaluation dataset as a collection of 5500 drum recordings spanning the E-GMD, MDB-Drums, Synthetic, Freesound Loops, and MusDB datasets. However, we erroneously omitted MusDB results (500 recordings) for round-trip reconstruction from Table 3. The missing reconstruction results for MusDB are as follows:

ModelMSL
D-Composer1.63
IDM1.67

Drum Class Counts

In Table 1 of the paper, we provide our chosen drum vocabulary with incorrect class counts; the correct class counts are as follows:

MIDIClassCount
36kick9007
37rim220
38snare6643
39clap1196
42hat4331
43tom4191
49cymbal9845
54tambourine350
56cowbell129
67perc4479
70shaker759
71other3708
Total44858