Datasets:

vocsim
/

avian-perception-benchmark

dataset_info:
  features:
    - name: index
      dtype: int64
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: subset
      dtype: string
    - name: speaker
      dtype: string
    - name: label
      dtype: string
    - name: original_name
      dtype: string
  splits:
    - name: train
      num_bytes: 24343526
      num_examples: 887
  download_size: 22452898
  dataset_size: 24343526
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: cc-by-4.0
tags:
  - audio
  - animal-vocalization
  - birdsong
  - zebra-finch
  - perceptual-similarity
  - benchmark
  - zero-shot
  - vocsim
  - avian-perceptual-judgment
  - audio-perceptual-judgment
size_categories:
  - n<1K
pretty_name: VocSim — Avian Perception Alignment

VocSim — Avian Perception Alignment

A companion dataset for the VocSim benchmark that tests whether neural audio embeddings align with biological perceptual judgments. It packages zebra finch (Taeniopygia guttata) song-syllable recordings together with the behavioral probe and triplet results from Zandberg et al. (2024), so an embedding's pairwise distance matrix can be compared directly against the birds' perceptual decisions.

Basha, M., Zai, A. T., Stoll, S., & Hahnloser, R. H. R. VocSim: A Training-free Benchmark for Zero-shot Content Identity in Single-source Audio. ICML 2026. arXiv:2512.10120

How to use it

Extract features from each syllable with the audio model you want to evaluate.
Compute pairwise distances between embeddings.
Score the distances against the behavioral judgments in probes.csv (probe trials) and triplets.csv (triplet trials).

The reference implementation lives at github.com/vocsim/benchmark — see reproducibility/scripts/avian_perception.py and reproducibility/configs/avian_paper.yaml.

Bundled files:

Hugging Face Dataset with the audio + metadata.
probes.csv — probe-trial results (sound_id, left, right, decision, …), filtered to rows whose audio is present.
triplets.csv — triplet-trial results (Anchor, Positive, Negative, diff, …), filtered the same way.
missing_audio_files.txt (when applicable) — original IDs without matching audio.

Schema

{
  "audio": {"array": np.ndarray, "sampling_rate": 16000},
  "subset": "avian_perception",
  "index": 42,
  "speaker": "ZF_M_123",                     # bird ID
  "label": "ZF_M_123",                       # set to speaker for this dataset
  "original_name": "ZF_M_123_syllable_A.wav" # identifier used in the CSVs
}

Quick start

from datasets import load_dataset

ds = load_dataset("vocsim/avian-perception-benchmark", split="train")
print(ds[0])

Source data

The recordings and behavioral results are from Zandberg et al. (2024). Please cite both that work and the VocSim paper if you use this dataset.

Citation

@inproceedings{basha2026vocsim,
  title     = {VocSim: A Training-free Benchmark for Zero-shot Content Identity in Single-source Audio},
  author    = {Basha, Maris and Zai, Anja T. and Stoll, Sabine and Hahnloser, Richard H. R.},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  year      = {2026},
  doi       = {10.48550/arXiv.2512.10120}
}

@article{zandberg2024bird,
  author    = {Zandberg, Lies and Morfi, Veronica and George, Julia M. and Clayton, David F. and Stowell, Dan and Lachlan, Robert F.},
  title     = {Bird song comparison using deep learning trained from avian perceptual judgments},
  journal   = {PLoS Computational Biology},
  volume    = {20},
  number    = {8},
  pages     = {e1012329},
  year      = {2024},
  doi       = {10.1371/journal.pcbi.1012329},
  publisher = {Public Library of Science}
}