README.md

metadata

license: cc-by-4.0
size_categories:
  - 100K<n<1M
pretty_name: VocSim
tags:
  - audio
  - audio-similarity
  - zero-shot-learning
  - representation-learning
  - embedding-evaluation
  - unsupervised-learning
  - speech
  - environmental-sounds
  - animal-vocalizations
  - benchmark
paperswithcode_id: audiosim
dataset_info:
  features:
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: subset
      dtype: string
    - name: speaker
      dtype: string
    - name: label
      dtype: string
  splits:
    - name: train
      num_bytes: 5452179735
      num_examples: 114641
  download_size: 5500616162
  dataset_size: 5452179735
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

VocSim — Public Benchmark

The public split of VocSim, a training-free benchmark for zero-shot content identity in single-source audio embeddings. VocSim probes the intrinsic geometric quality of frozen audio representations: do acoustically variable instances of the same content land near each other in embedding space, without any task-specific training?

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

What's here

• 114,641 clips across 15 public subsets, drawn from 19 source corpora.
• Domains: human speech (phones, words, utterances), animal vocalizations (birdsong, otter calls), environmental sounds.
• Conditions: clean to noisy, sub-100ms to multi-second, few to thousands of classes per subset.
• All audio standardized to 16 kHz mono.
• Single-source only — no overlapping speakers or simultaneous sources — so evaluation isolates content representation from source separation.

Four additional blind out-of-distribution subsets (low-resource speech in Shipibo-Conibo and Chintang) are held out for server-side evaluation via the leaderboard.

Schema

{
  "audio":  {"array": np.ndarray, "sampling_rate": 16000},
  "subset": "HW1",        # source-corpus tag (see paper for the full list)
  "speaker": "spk_042",   # speaker / animal / source ID, or "N/A"
  "label":  "hello",      # ground-truth class for similarity
}

Quick start

from datasets import load_dataset

ds = load_dataset("vocsim/public", split="train")
print(ds[0])

For end-to-end evaluation (feature extraction, distance computation, P@k / GSR), use the reference pipeline at github.com/vocsim/benchmark.

Companion datasets

Dataset	Purpose
vocsim/avian-perception-benchmark	Alignment of embeddings with zebra-finch perceptual judgments
vocsim/mouse-strain-classification-benchmark	C57 vs DBA USV classification
vocsim/mouse-identity-classification-benchmark	Individual-mouse identification from USVs

Licensing

Aggregation and metadata are released under CC BY 4.0. Each source corpus retains its original license; see Appendix A.1.1 of the paper for a per-source breakdown.

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}
}