A-COAT-2k

Audio Compositional Object Algebra Test — 2,000 zero-shot audio quadruples for evaluating whether audio encoders represent multi-source scenes compositionally. No training required.

Companion dataset to the ICASSP 2026 paper Evaluating Compositional Structure in Audio Representations. See also the trained-head benchmark chuyangchenn/a-tre-10k.

Quick start

from datasets import load_dataset

ds = load_dataset("chuyangchenn/a-coat-2k", split="test")
ex = ds[0]
A = ex["A"].get_all_samples().data   # torch.Tensor, shape (1, 320000)
B = ex["B"].get_all_samples().data   # B = A ∪ T
C = ex["C"].get_all_samples().data
D = ex["D"].get_all_samples().data   # D = C ∪ T
metadata = ex["metadata"]            # {"A": [...], "C": [...], "T": [...]}

What's a quadruple?

Each row is a 4-tuple (A, B, C, D) where B = A ∪ T and D = C ∪ T — i.e. the same transformation set T is applied to two different base scenes. The benchmark score for an encoder f is the average over quadruples of:

A-COAT(A,B,C,D) = cos(f(B) − f(A), f(D) − f(C))

Score 1 = adding T shifts the embedding by the same vector regardless of base scene (perfect compositionality). Random encoders score ≈ 0.

Dataset structure

Field	Type	Description
A,B,C,D	Audio(sampling_rate=32000)	Waveforms, each (1, 320000) mono 32 kHz.
metadata	dict[str, list[dict]]	Source attributes per role: A, C, T.

Each source has four discrete attributes (K = 8 classes per attribute):

• timbre — t1–t8: eight DX7 FM synth patches
• pitch — p1–p8: MIDI 36–84, linearly binned
• rate — r1–r8: 0.2–3.0 Hz, log-binned repetition rate
• amplitude — a1–a8: −26 to 0 dB, linearly binned

A and C each contain 1 source. T contains 1–3 sources (varies per quadruple).

Splits

Split	# quadruples
test	2,000

(No train/val — A-COAT is a zero-shot benchmark.)

Citation

@inproceedings{chen2026audiocomp,
  title         = {Evaluating Compositional Structure in Audio Representations},
  author        = {Chen, Chuyang and Steers, Bea and McFee, Brian and Bello, Juan Pablo},
  booktitle     = {IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year          = {2026},
  eprint        = {2603.13685},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD}
}

License

CC BY 4.0 — free use with attribution.