VoiceTrace-Bench

VoiceTrace is a benchmark and unified framework for who-said-what speech retrieval: given a natural-language query about a speaker's identity or what they said, retrieve the matching audio document. Unlike conventional speaker verification or diarization benchmarks, VoiceTrace evaluates retrieval jointly over who is speaking and what is being said, across both single-speaker and multi-speaker conversational recordings.

This repository hosts the VoiceTrace-Bench evaluation dataset, which has two splits:

• single (1,510 pairs) — single-speaker query/document audio pairs, drawn from the VoxCeleb1 and VoxCeleb2 test sets.
• multi (852 pairs) — multi-speaker conversational recordings, drawn from VoxConverse and Seamless Interaction.

Each example pairs a query (natural-language question plus a query audio clip) against the document audio clip that answers it.

Dataset Structure

Each split is stored as data.parquet alongside its audio files. Both splits share an identical schema, so they load as one dataset with two splits. Audio paths are relative to the split directory, and the ground-truth document for a query is the one on the same row.

Field	Type	Description
query_text	string	Natural-language query about the speaker or what they said
query_audio_sample_rate	int64	Sample rate of the query audio, in Hz
query_audio_duration	float64	Duration of the query audio, in seconds
document_audio_sample_rate	int64	Sample rate of the document audio, in Hz
document_audio_duration	float64	Duration of the document audio, in seconds
query_audio_path	string	Path to the query audio clip, relative to the split directory
document_audio_path	string	Path to the document audio clip, relative to the split directory

An example record:

{
  "query_text": "How does this speaker justify the high salaries and performances of professional athletes?",
  "query_audio_sample_rate": 16000,
  "query_audio_duration": 1.0,
  "document_audio_sample_rate": 16000,
  "document_audio_duration": 34.0869375,
  "query_audio_path": "seamless_interaction_samples/audio/row_000000_query.wav",
  "document_audio_path": "seamless_interaction_samples/audio/row_000000_document.wav"
}

Usage

Download the dataset with the Hugging Face CLI:

hf download cara-ai/VoiceTrace-Bench --local-dir benchmark

Or load a split directly with datasets:

from datasets import load_dataset

single = load_dataset("cara-ai/VoiceTrace-Bench", split="single")
multi = load_dataset("cara-ai/VoiceTrace-Bench", split="multi")

Then evaluate a baseline speaker retrieval model against it using the VoiceTrace evaluation script:

git clone git@github.com:caml-labs/VoiceTrace.git
cd VoiceTrace
conda create -n eval python=3.12
conda activate eval
pip install -r requirements.txt
python evaluate_baseline.py --single-dataset-path benchmark/single --multi-dataset-path benchmark/multi

Citation

If you find this work useful, please consider citing:

@misc{yee2026voicetrace,
      title={VoiceTrace: A Benchmark and Retrieval Framework for Who-Said-What Speech Retrieval}, 
      author={Aaron Yee and Fengjie Lu and Jiarui Hai and Chenang Jiang and Helin Wang and Siwei Tu and Weitao You and Lingyun Sun},
      year={2026},
      eprint={2609.18521},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2609.18521}, 
}

License

This dataset is licensed under CC-BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International).

Acknowledgements

• VoxCeleb / VoxCeleb2 for providing large-scale benchmark datasets for speaker-related research.
• VoxConverse for providing multi-speaker conversational recordings with diarization annotations.
• Seamless Interaction for open-source resources that support speech and multimodal interaction research.