VocSim: A Training-Free Benchmark for Content Identity in Single-Source Audio Embeddings

VocSim evaluates how well neural audio embeddings generalize for zero-shot audio similarity. It tests recognizing fine-grained acoustic similarity without specific similarity training.

---

Key Features

• Diverse Sources: Human speech (phones, words, utterances), birdsong, otter calls, environmental sounds.
• Varied Conditions: Spans clean to noisy recordings, short (<100ms) to long durations, few to many classes per subset.
• Standardized: All audio is 16kHz mono.

---

Data Format

{
  'audio': {'array': array([...], dtype=float32), 'sampling_rate': 16000},
  'subset': 'HW1',      # Origin identifier
  'speaker': 'spk_id',  # Speaker/Animal/Source ID or 'N/A'
  'label': 'hello'      # Ground truth class for similarity
}

Train split: 114,641 public examples from 15 subsets for evaluation.

Blind Test Sets: 4 additional subsets held out privately.

Citation

@inproceedings{vocsim_authors_2025,
  title={VocSim: A Training-Free Benchmark for Content Identity in Single-Source Audio Embeddings},
  author={Anonymous Authors},
  booktitle={Conference/Journal},
  year={2025},
  url={[Link to paper upon DOI]}
}

License

CC BY 4.0 - Creative Commons Attribution 4.0 International.