README.md

---
pretty_name: 'SPARC: Spoken Profile Authorship Recognition Corpus'
language:
- en
license: cc-by-nc-4.0
license_name: sparc-mixed
task_categories:
- text-classification
tags:
- authorship-verification
- speaker-verification
- privacy
- speech-transcripts
- podcasts
size_categories:
- 1M<n<10M
viewer: false
extra_gated_prompt: >-
  SPARC is released for privacy research only. Do not use it to re-identify or
  profile individuals, or for forensic, law-enforcement, employment,
  eligibility, or other high-stakes decisions. Sex, age, and accent labels are
  model-inferred, not verified identities. The SPoRC-derived portion is subject
  to the SPoRC data-use terms (research and education only).
extra_gated_fields:
  Affiliation: text
  Intended use: text
  I agree to use SPARC only for research on transcript privacy and its defenses: checkbox
  I agree to the SPoRC data-use terms for the SPoRC-derived file: checkbox
---

<p align="center">
  <img src="logo.png" alt="SPARC logo" width="300">
</p>

# SPARC: Spoken Profile Authorship Recognition Corpus

SPARC is a benchmark for measuring how much **speaker identity** and **demographic attributes** (sex, age, accent) leak from transcribed speech. It contains quality-filtered, single-speaker podcast transcripts from two corpora, NaturalVoices and SPoRC, with speaker-level attribute labels.

**This repository contains the speaker-level data only.** The code, the verification-trial generation used in the paper (identity, sex, age, and accent trials; probe length P ∈ {50, 200} words; enrollment size L ∈ {5, 20} utterances), and the evaluation scripts are on GitHub: [link](https://github.com/bowenyi-pierre/SPARC)

## Files

| File | Source | Speakers | Episodes | Utterances | Words | Size |
|---|---|---:|---:|---:|---:|---:|
| `naturalvoices_speakers.json` | NaturalVoices (CC-BY 4.0 episodes) | 2,344 | 3,500 | 779,494 | 13.7M | 87 MB |
| `sporc_speakers.json` | SPoRC (single-speaker subset) | 10,165 | 15,555 | 795,591 | 26.5M | 153 MB |

Both files share the same structure.

## Structure

Each file is a JSON object keyed by speaker ID. Each speaker maps to one key per episode (that speaker's utterances in that episode) plus five attribute keys:

```json
{
  "n00030": {
    "MSP-PODCAST_0001": [
      "Hey, you are listening to Nothing Sounds Better Podcast.",
      "..."
    ],
    "age": 28.78,
    "gender": "male",
    "accent": "North America",
    "accent_probs": {"Eastern Asia": 0.0001, "English": 0.001, "...": "...", "North America": 0.9688, "Welsh": 0.0},
    "accent_num_utterances": 16
  }
}
```

| Key | Type | Description |
|---|---|---|
| `<episode_id>` | `list[str]` | The speaker's retained utterances in that episode, in temporal order. Every key that is not one of the five attributes below is an episode. |
| `age` | `float` | Predicted age in years, averaged over utterances. |
| `gender` | `str` | Predicted sex: `male` or `female`. NaturalVoices also contains `child` for 5 speakers. |
| `accent` | `str` | Argmax of `accent_probs`. |
| `accent_probs` | `dict[str, float]` | Speaker-level distribution over 16 English dialect classes: Eastern Asia, English, Germanic, Irish, North America, Northern Irish, Oceania, Other, Romance, Scottish, Semitic, Slavic, South African, South-eastern Asia, Southern Asia, Welsh. |
| `accent_num_utterances` | `int` | Number of utterance-level predictions averaged into `accent_probs`. |

### Identifiers

| | NaturalVoices | SPoRC |
|---|---|---|
| Speaker ID | `n` + original NaturalVoices speaker ID, zero-padded (`n00030` = speaker 30) | `s00000`–`s10164`, pseudonymous; speaker names are not released |
| Episode ID | `MSP-PODCAST_####`, from NaturalVoices | `ep_######`, internal episode index |
| Episodes | May contain several speakers (1,699 of 3,500) | Single-speaker |

## Loading

The files are nested JSON (one object per speaker), so `load_dataset` and the dataset viewer do not apply. Load them directly:

```python
import json
from huggingface_hub import hf_hub_download

ATTRS = {"age", "gender", "accent", "accent_probs", "accent_num_utterances"}

path = hf_hub_download("timtsapras23/SPARC", "naturalvoices_speakers.json", repo_type="dataset")
with open(path) as f:
    data = json.load(f)

for speaker_id, info in data.items():
    episodes = {k: v for k, v in info.items() if k not in ATTRS}  
    labels = {k: info[k] for k in ATTRS}
```

Use the GitHub scripts to build the verification trials from these files.

## Construction

SPARC is built from two English podcast corpora: [NaturalVoices](https://arxiv.org/abs/2511.00256) (Du et al., 2025) and [SPoRC](https://arxiv.org/abs/2411.07892) (Litterer et al., 2025). Speaker labels are either taken from the original data or inferred with speech models:

- **Age and sex.** Taken from the original metadata for NaturalVoices; inferred with [Vox-Profile](https://arxiv.org/abs/2505.14648) (Feng et al., 2025b) for SPoRC.
- **Accent.** Inferred with [Voxlect](https://arxiv.org/abs/2508.01691) (Feng et al., 2025a) for both corpora.

Details are in the paper.

## Statistics

| | NaturalVoices | SPoRC |
|---|---:|---:|
| Speakers | 2,344 | 10,165 |
| Unique episodes | 3,500 | 15,555 |
| Speaker–episode pairs | 6,766 | 15,555 |
| Utterances | 779,494 | 795,591 |
| Avg. utterances / speaker | 332.6 | 78.3 |
| Mean age (years) | 41.5 | 37.9 |
| Male / Female / Child (%) | 52.4 / 47.4 / 0.2 | 59.3 / 40.7 / 0 |
| Age < 30 / 30–60 / 60+ (%) | 15.7 / 76.2 / 8.1 | 36.3 / 53.1 / 10.6 |
| North American accent (%) | 78.2 | 77.3 |


## Intended Use

SPARC is intended for research on transcript privacy: measuring identity and attribute leakage from transcribed speech, and developing and evaluating text-level anonymization and other defenses.

**Out of scope:** re-identifying, profiling, or making decisions about individuals, including forensic, law-enforcement, employment, eligibility, or other high-stakes uses.

## License

- `naturalvoices_speakers.json`: derived from NaturalVoices episodes released under CC-BY 4.0; redistributed under **CC-BY 4.0**. Please credit NaturalVoices (Du et al., 2025).
- `sporc_speakers.json`: derived from SPoRC; use is subject to the [SPoRC data-use terms](https://huggingface.co/datasets/blitt/SPoRC) (research and education only, no commercial use). Please credit SPoRC (Litterer et al., 2025).
- Accent labels are produced with Voxlect; reuse should comply with Voxlect's Responsible AI License.


## Citation
 
```bibtex
@misc{tsaprazlis2026sparcspokenprofileauthorship,
      title={SPARC: Spoken Profile Authorship Recognition Corpus}, 
      author={Efthymios Tsaprazlis and Bowen Yi and Tiantian Feng and Junghwan Kim and David Jurgens and Shrikanth Narayanan},
      year={2026},
      eprint={2610.25265},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.25265}, 
}
```

Please also cite the source corpora and labeling models:

- Du et al. 2025. *NaturalVoices: A Large-Scale, Spontaneous and Emotional Podcast Dataset for Voice Conversion.* arXiv:2511.00256.
- Litterer, Jurgens, and Card. 2025. *Mapping the Podcast Ecosystem with the Structured Podcast Research Corpus.* ACL 2025.
- Feng et al. 2026. *Vox-Profile: A Speech Foundation Model Benchmark for Characterizing Diverse Speaker Traits and Speech Properties.* The Journal of Data-centric Machine Learning Research (DMLR).
- Feng et al. 2026. *Voxlect: A Speech Foundation Model Benchmark for Modeling Dialects and Regional Languages Around the Globe.* KDD 2026.