README.md

---
license: cc-by-nc-4.0
language:
- en
pretty_name: VIPBench
size_categories:
- 100K<n<1M
tags:
- speaker-recognition
- voice-identity
- voice-cloning
- human-perception
- benchmark
- audio
task_categories:
- audio-classification
- other
configs:
- config_name: judgments
  data_files: data/participant_responses.csv
  default: true
- config_name: stimuli
  data_files: data/stimuli.csv
- config_name: speakers
  data_files: data/speakers.csv
- config_name: morph_metadata
  data_files: data/stimuli_interpol.csv
---

# VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning

VIPBench is a benchmark of **124,876 same/different identity judgments** from **1,290 English-speaking listeners** on **9,800 voice pairs** spanning **100 demographically-stratified speakers**. Pairs cover three stimulus families: real recordings, AI voice clones generated by a state-of-the-art TTS system, and continuously morphed voices.

The benchmark evaluates whether speaker-embedding and speech-representation models align with human voice-identity perception, providing a perceptual evaluation target distinct from metadata-label speaker identification.

> **Anonymized release for NeurIPS 2026 Evaluations & Datasets Track double-blind review.**
> Author identities and permanent URLs will be added at camera-ready.

---

## Dataset summary

| Item | Count |
|---|---|
| Speakers | 100 (50 M / 50 F, 5 sociophonetic groups, 2 age brackets) |
| Reference audio clips | 100 (one per speaker) |
| Comparison audio clips | 9,800 (98 per speaker) |
| Voice pairs | 9,800 |
| Listener judgments | 124,876 |
| Listeners | 1,290 |
| Median judgments per pair | 10 (range 8 to 92) |
| Stimulus types | 6 (real same/different, AI clones, voice morphs) |
| Pre-extracted speaker embeddings | 10 models |
| Per-layer SSL embeddings | 5 models |

## Supported tasks

The benchmark defines four evaluation tasks:

1. **Predict listener agreement rate** (continuous regression). Predict `P(same)` per pair. Metric: Pearson r, R^2 against the human consensus, bounded by the Spearman-Brown noise ceiling rho_SB = 0.705.
2. **Human-aligned binary verification**. Classify pairs against the human majority vote. Metrics: AUC (ranking) and Platt-calibrated ECE (calibration).
3. **Representational similarity (RSA)**. Spearman correlation between human and model representational dissimilarity matrices, with a Mantel permutation test.
4. **Real-to-synthetic transfer**. Whether a predictor fit on real-speech pairs still works on voice clones and morphs.

A 10-fold gender-balanced speaker-level cross-validation protocol prevents speaker leakage.

## Dataset structure

```
data/
  speakers.csv                   # 100 rows: speaker id, name, group, gender, age
  stimuli.csv                    # 9,800 rows: per-pair aggregates (P(same), votes, type)
  participant_responses.csv      # 124,876 rows: per-judgment records
  stimuli_interpol.csv           # 8,100 rows: morph-trajectory metadata for Type 6
  audio/
    reference/                   # 100 *R.wav (16 kHz mono)
    comparison/                  # 9,800 *.wav
  embeddings/
    rawnet3.npz, ecapa_tdnn.npz, titanet.npz, xvector.npz, resemblyzer.npz,
    wav2vec2.npz, hubert.npz, wavlm.npz, xlsr.npz, whisper.npz
    layers/                      # per-layer (mean-pooled) for SSL models
      wav2vec2.npz, hubert.npz, wavlm.npz, xlsr.npz, whisper.npz
samples/                         # 5-speaker quick-look subset (~150 MB)
code/                            # 10 extraction scripts + analysis notebook + reproduce.sh
docs/                            # annotation protocol, schemas, model table, reproduction
```

For column-level dictionaries see `docs/data_dictionary.md`. For the six stimulus types see `docs/stimulus_types.md`. For the listening-study protocol see `docs/annotation_protocol.md`.

### Embedding format

Each `.npz` is a key-value store keyed by audio basename without the `.wav` extension (e.g., `M01R`, `1_F01`, `4_M12_M15B`). Values are numpy arrays of shape `(embedding_dim,)` for the 10 main embeddings and `(num_layers, embedding_dim)` for the per-layer bundles. The 9,900 keys cover 100 references plus 9,800 comparisons.

### Pairing reference and comparison

Each row of `data/stimuli.csv` represents one voice pair. The `reference` column gives the reference speaker ID (e.g., `M01`) and the `id` column gives the stimulus identifier of the comparison clip (e.g., `1_M01`, `4_M12_M15B`). The pairing rule is:

| You want | Reference clip | Comparison clip |
|---|---|---|
| Audio file | `data/audio/reference/{row.reference}R.wav` | `data/audio/comparison/{row.id}.wav` |
| Embedding key | `{row.reference}R` (e.g., `M01R`) | `{row.id}` (e.g., `1_M01`) |

Cosine-similarity scoring against `P(same)`:

```python
import numpy as np, pandas as pd
from sklearn.metrics.pairwise import cosine_similarity

stim = pd.read_csv('data/stimuli.csv')
emb  = dict(np.load('data/embeddings/ecapa_tdnn.npz'))

stim['cos'] = stim.apply(
    lambda r: cosine_similarity(
        emb[f'{r.reference}R'].reshape(1, -1),
        emb[r.id].reshape(1, -1)
    )[0, 0],
    axis=1,
)
stim['p_same'] = stim['same_vote'] / stim['num_response']
print(stim[['cos', 'p_same']].corr())   # Pearson r against listener consensus
```

## Quick start

```bash
pip install -r requirements.txt
cd code && bash reproduce.sh        # ~10 min from cached embeddings
```

To re-extract embeddings from the audio (~24 CPU-hours plus ~1 GPU-hour for Whisper), see `docs/reproduction.md`.

## Source data and collection

- **Speakers.** 100 English-speaking US celebrities stratified across 5 sociophonetic groups (1 = New York City English, 2 = Southern American English, 3 = African American English, 4 = Latino English, 5 = Asian American English) x 2 genders x 2 age brackets (1 = under 45, 2 = 55 or older), 5 speakers per cell.
- **Reference audio.** Clips selected from publicly available recordings (interviews, podcasts).
- **Voice clones.** Generated with Cartesia (a state-of-the-art TTS system) seeded from a natural source clip of the speaker being cloned. The variant letter in the stimulus ID identifies the seed: a Type 3 clone shares its seed clip with the comparison clip of the matched Type 2 pair, and a Type 5 clone shares its seed with the matched Type 4 pair (e.g., the clone in `3_F01B` is seeded from the same F01B source clip used as the comparison in `2_F01B`).
- **Voice morphs.** For each of the 100 reference speakers, the latent voice representation of the reference speaker is interpolated toward each of 4 within-group comparison speakers (matched on sociophonetic group, age group, and gender), at 2 distinct recordings per comparison speaker, sampled at 10 morph scales between 0 and 1, plus 1 anchor at scale 1. This yields 4 x 2 x 10 + 1 = 81 Type-6 stimuli per reference speaker (8,100 total). Generated using the voice-morphing feature of the same Cartesia TTS system.
- **Listeners.** 1,290 adult English-speaking participants recruited via the Centaur AI platform under an IRB-approved protocol. Consent followed the platform's standard pipeline.

Each pair received at least 8 judgments; real-speech pairs (Types 1, 2, 4) received more coverage than synthetic pairs to give tighter consensus estimates on the real-speech reference distribution.

## Considerations for use

### Personally identifying information

The dataset names public-figure speakers because the celebrity-stratified design is integral to the benchmark and source recordings are already public. Listener identifiers in `participant_responses.csv` are pseudonymized integers tied to no external account.

### Biases and limitations

- English-speaking listener pool, US-dialect speakers. Cross-language perception is not measured.
- 100 speakers limits statistical power for some subgroup contrasts (20 speakers per sociophonetic group).
- Studio-quality audio. In-the-wild conditions (noise, codec compression, telephony) are not represented.
- The operational target is a population consensus, appropriate for ambiguous stimuli where any absolute identity label would itself be probabilistic.

### Responsible use

The benchmark measures model-human alignment at the evaluation level. We do not release clone-generation recipes or adversarial training targets. Voice-cloning systems that better align with human perception could inform adversarial use; the same alignment knowledge also strengthens defenses (perception-aligned identity models can flag clones that metadata-based verification accepts).

## License

- **Dataset (audio, judgments, metadata, embeddings):** Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0). See `LICENSE`.
- **Code (scripts, notebook):** MIT License. See `LICENSE-CODE`.
- **Pretrained model weights** (loaded by extraction scripts): each baseline retains its original license; see `docs/model_table.md`.

Commercial use of the audio, judgments, or derived embeddings is not permitted under this license.

## Citation

To be filled in at camera-ready.

```
@inproceedings{vipbench2026,
  title  = {VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning},
  author = {Anonymous},
  booktitle = {Advances in Neural Information Processing Systems Datasets and Benchmarks},
  year   = {2026},
  note   = {Anonymized for double-blind review.}
}
```

## Files

- `README.md` (this file): dataset card.
- `LICENSE`: CC-BY-NC 4.0 full text.
- `LICENSE-CODE`: MIT full text for scripts.
- `croissant.json`: MLCommons Croissant 1.0 metadata (core + Responsible AI fields).
- `DATASHEET.md`: Datasheet for Datasets (Gebru et al. 2021).
- `CHANGELOG.md`: version history.
- `CITATION.cff`: machine-readable citation.
- `requirements.txt`: pinned Python dependencies.
- `data/`, `samples/`, `code/`, `docs/`: see structure section above.