Drum Transcription

Automatic drum transcription (ADT) aims at extracting note onsets of percussion instrument events from a music recording. An accurate ADT system enables diverse applications in music education, music production, music search and recommendation, and computational musicology.

Limitations of existing ADT models

There are over 500 percussion instruments that exist around the world and new ones appear all the time. However, most existing ADT systems only focus on transcribing 3 instruments - bass drum, snare drum, and hi-hat. This is again due to the limited number and size of ADT datasets.

More recent approaches leveraging synthetic data and deep neural networks [41][5] expand their output vocabulary to transcribe up to 18 instruments. While it is a significant improvement, 10-20 classes are still far from the wide gamut of percussive instruments used in recorded music. For example, rare or non-western percussion sounds are usually considered out-of-vocabulary. Moreover, when transcribing different datasets, we often need to manually map the percussion instruments in a dataset to the limited output vocabulary of an existing ADT system with reduced granularity. It can also be challenging for ADT systems that utilize fully-supervised learning to generalize to different musical genres or diverse drum sounds. All of these can greatly limit the real-world applicability of ADT. However, further expanding the output vocabulary would typically require a large amount of labeled data per additional class, hence require a lot of annotation effort.

Few-shot ADT

Building on top of the success of few-shot classification among monophonic audio, Wang et al. [10] propose to apply the technique developed in the work of few-shot sound event detection [39] to polyphonic musical audio to perform few-shot ADT, achieving open-vocabulary ADT based on few labeled data. They frame the drum transcription task as a per-frame binary sound classification task, each targeting one instrument. With a few-shot model, pre-trained on percussion instruments from a large synthetic dataset [6], we can transcribe any percussion instrument of interest by providing a few examples of the target.

The proposed few-shot ADT framework is evaluated on multiple real-world ADT datasets with polyphonic accompaniment. The result shows that, given just a handful of selected examples at inference time, the few-shot approach can match and in some cases outperform a state-of-the-art supervised ADT model under a fixed vocabulary setting. At the same time, the few-shot model can successfully generalize to finer-grained or extended vocabularies unseen during training, a scenario where supervised approaches cannot operate at all.