A-TRE-10k

Audio Tree Reconstruction Error benchmark — 10,000 synthetic audio scenes for evaluating whether audio encoders represent multi-source scenes compositionally.

Companion dataset to the ICASSP 2026 paper Evaluating Compositional Structure in Audio Representations. See also the zero-shot benchmark chuyangchenn/a-coat-2k.

Quick start

from datasets import load_dataset

ds = load_dataset("chuyangchenn/a-tre-10k", split="train")  # or "val", "test"
ex = ds[0]
samples = ex["audio"].get_all_samples()
waveform = samples.data           # torch.Tensor, shape (1, 320000)
sr = samples.sample_rate          # 32000
metadata = ex["metadata"]         # list of {timbre_label, pitch_label, rate_label, amplitude_label}

Streaming (no local download):

ds = load_dataset("chuyangchenn/a-tre-10k", split="test", streaming=True)
for ex in ds.take(5):
    print(ex["audio"].get_all_samples().data.shape)

Dataset structure

Each row is one 10-second 32 kHz mono audio scene plus its source-attribute metadata.

Field	Type	Description
audio	Audio(sampling_rate=32000)	Waveform, shape (1, 320000) — peak-normalised mono.
metadata	list[dict]	One entry per source: {timbre_label, pitch_label, rate_label, amplitude_label}.

A scene contains N ∈ {1, 2, 3, 4} independent sources, each described by 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

Splits

Split	# scenes
train	8,000
val	1,000
test	1,000

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.