kac_drumset: A Dataset Generator for Arbitrarily Shaped Drums

This publication documents the various datasets generated using the kac_drumset codebase. The aims of kac_drumset is to provide a robust framework for the generation and analysis of arbitrarily shaped drums. The source code for this project is available here: https://github.com/lewiswolf/kac_drumset.

Background

Arbitrarily shaped drums are a strange family of percussion instruments and a wholly meta-physical construction in this contemporary setting. These percussive instruments possess a number of interesting musical characteristics resulting from their particular geometric designs. As it currently stands, these instruments remain largely unexplored throughout musical practice, as they were originally devised as a collection of hypothetical mathematical objects. These datasets serve to sonify these objects so as to explore these conceptual constructions in the audio domain.

Usage

To use these datasets, first install kac_drumset:

pip install "git+https://github.com/lewiswolf/kac_drumset.git#egg=kac_drumset"

And then in python:

from kac_drumset import (
# methods
loadDataset,
transformDataset,
# classes
TorchDataset,
)

dataset: TorchDataset = transformDataset(
# load a dataset (any folder containing a metadata.json)
loadDataset('absolute/path/to/data'),
# alter the dataset representation, either as an end2end, fft or mel.
{'output_type': 'end2end'},
)

# use the dataset
for i in range(dataset.__len__()):
x, y = dataset.__getitem__(i)
...

For more details on using kac_drumset, see the project's documentation.

2000 Convex Polygonal Drums of Varying Size

Each sample in this dataset corresponds to a randomly generated convex polygon. The audio for each sample was generated using a two-dimensional physical model of a drum. Each sample is one second long and decays linearly.

Contained in this dataset are ten different sizes of drums - 0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.5, 0.6 - each of which is a measure of the longest vertex of each drum in meters. There are 40 different drums sampled for each size. Each drum is sampled five times, first by being struck in the geometric centroid, and then by being struck four more times in random locations. This dataset is labelled with the vertices of each polygon, normalised to the unit interval, and the strike location of each sample.

The audio is sampled at 48khz, and the default representation is raw audio. Each sample is stored in the metadata.json, alongside being made available audibly as a 24-bit .wav and graphically as a .png.

5000 Circular Drums of Varying Size

Each sample in this dataset corresponds to a randomly generated circular drum. The audio for each sample was generated using additive synthesis, inferred using a closed form solution to the two dimensional wave equation. Each sample is one second long and decays exponentially.

Contained in this dataset are 1000 different drums, each determined by a randomly generated size (0.1, 2.0) in meters. Each drum is sampled five times, first being struck in the geometric centroid, and then by being struck four more times in random locations. This dataset is labelled with the size of each drum and the strike location of each sample.

The audio is sampled at 48khz, and the default representation is raw audio. Each sample is stored in the metadata.json, alongside being made available audibly as a 24-bit .wav and graphically as a .png.

5000 Rectangular Drums of Varying Dimension

Each sample in this dataset corresponds to a randomly generated rectangular drum. The audio for each sample was generated using additive synthesis, inferred using a closed form solution to the two dimensional wave equation. Each sample is one second long and decays exponentially.

Contained in this dataset are 1000 different drums, each determined by a randomly generated size (0.1, 2.0) in meters and aspect ratio (0.25, 4.0). Each drum is sampled five times, first being struck in the geometric centroid, and then by being struck four more times in random locations. This dataset is labelled with the size and aspect ratio of each drum, and the strike location of each sample.

The audio is sampled at 48khz, and the default representation is raw audio. Each sample is stored in the metadata.json, alongside being made available audibly as a 24-bit .wav and graphically as a .png.