# Datasheet for VIPBench

Following the structure of Gebru et al. 2021, "Datasheets for Datasets" (Communications of the ACM 64(12)).

## Motivation

**For what purpose was the dataset created?**
To benchmark whether speaker-embedding and speech-representation models align with human voice-identity perception. Existing speaker benchmarks evaluate against a metadata speaker label (who produced the recording); this dataset measures whether listeners hear two voices as the same speaker, which can diverge from the metadata label especially under voice cloning and voice morphing.

**Who created the dataset and on behalf of which entity?**
Withheld for double-blind review. Will be filled in at camera-ready.

**Who funded the creation of the dataset?**
Withheld for double-blind review.

## Composition

**What do the instances represent?**
Each instance is a voice pair (one reference audio clip plus one comparison audio clip) annotated with same-or-different-speaker judgments collected from human listeners.

**How many instances are there?**
9,800 voice pairs, 124,876 listener judgments, 1,290 listeners, 100 speakers, 9,900 audio files (100 reference + 9,800 comparison).

**Does the dataset contain all possible instances or is it a sample?**
It is a designed sample. 100 speakers were selected to balance 5 sociophonetic groups (New York City English, Southern American English, African American English, Latino English, Asian American English) x 2 genders x 2 age brackets (under 45, and 55 or older), 5 speakers per cell. Pairs were sampled to cover six stimulus types: same recording (Type 1, 100 pairs), same speaker different recording (Type 2, 400 pairs), same speaker AI clone (Type 3, 400 pairs), different speaker different recording (Type 4, 400 pairs), different speaker AI clone (Type 5, 400 pairs), and continuously morphed pairs (Type 6, 8,100 pairs).

**What data does each instance consist of?**
- A reference audio clip (16 kHz mono WAV, ~6 seconds).
- A comparison audio clip (16 kHz mono WAV, similar duration).
- Per-pair aggregates (`stimuli.csv`): stimulus type, source speaker(s), morph scale (Type 6 only), counts of same/different votes, accuracy against the metadata label, derived `P(same)`.
- Per-judgment records (`participant_responses.csv`): listener ID, stimulus ID, binary same/different answer, listener-recognition flag, listener demographics (age band, gender, first language).

**Is there a label or target associated with each instance?**
The primary target is the human consensus `P(same)` per pair. A binary majority-vote label (`y = 1[P(same) > 0.5]`) is also supplied. The ground-truth metadata speaker label is included for reference.

**Is any information missing from individual instances?**
Listener demographics include age band (categorical), gender, and first-language flag. Detailed listener demographics or browser/device information are not included. The `know_speaker` field is missing for some early-trial responses; trials with the listener flagging recognition are still released and noted.

**Are relationships between individual instances made explicit?**
Yes. `stimuli.csv` indexes pairs by reference and comparison speaker IDs; `participant_responses.csv` indexes by listener ID and stimulus ID. The 9,900-key embedding files use the audio basename as key, allowing exact join with `stimuli.csv.id`.

**Are there recommended data splits?**
Speaker-level cross-validation (10 folds, gender-balanced 5M+5F per fold) is recommended to avoid speaker leakage. The benchmark protocol uses speaker-level GroupKFold throughout.

**Are there any errors, sources of noise, or redundancies?**
Identity perception is intrinsically noisy: split-half Spearman-Brown reliability of `P(same)` is rho_SB = 0.705, bounding any model's correlation against the observed target. Voice clones and mid-morph stimuli are designed to be ambiguous, so consensus is itself probabilistic.

**Is the dataset self-contained or does it link to external resources?**
Self-contained for evaluation: audio, judgments, and pre-extracted embeddings are bundled. Pretrained model weights are downloaded from Hugging Face / model hubs by the extraction scripts; their checkpoint identifiers are fixed in `docs/model_table.md`.

**Does the dataset contain confidential data?**
No. Listener identifiers are pseudonymized integers tied to no external account. Speaker names are public figures whose recordings are publicly available.

**Does the dataset contain offensive, insulting, or threatening content?**
No. Stimuli are conversational speech.

## Collection process

**How was the data acquired?**
- Reference audio: selected from publicly available recordings (interviews, podcasts) of 100 US celebrities.
- Comparison audio: a mix of additional natural clips of the same or different speakers (Types 2 and 4); AI voice clones generated with the Cartesia TTS system, seeded from a natural source clip of the speaker being cloned (Types 3 and 5; the variant letter in the stimulus ID identifies the seed, which is the same source clip used as the comparison in the matched Type 2 or Type 4 pair); and continuously morphed pairs (Type 6) generated using the voice-morphing feature of the same Cartesia system. For each reference speaker, morphs interpolate the latent voice representation 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: 4 x 2 x 10 + 1 = 81 stimuli per reference speaker, 8,100 total.
- Judgments: 1,290 adult English-speaking listeners recruited via the Centaur AI platform. Each pair received at least 8 judgments.

**What mechanisms or procedures were used to collect the data?**
Listeners heard each pair as a single audio clip (reference + 1 second silence + short beep + comparison) and answered (a) whether the two clips came from the same speaker (binary) and (b) optionally identified which of four within-group celebrities or "I don't know" they recognized in the reference.

**Who was involved in the data collection process?**
Listeners were paid via the Centaur AI platform per the platform's standard rate, which meets minimum-wage requirements in the country of data collection. The annotation protocol was approved by an Institutional Review Board.

**Over what time frame was the data collected?**
Stimulus creation and listening study were conducted between 2025 and 2026.

**Were any ethical review processes conducted?**
Yes. Both the clone generation (from publicly available celebrity speech) and the human-annotation study were covered under IRB review. Participant consent followed the platform's standard consent pipeline.

## Preprocessing, cleaning, labeling

**Was preprocessing or cleaning done?**
- Audio was resampled to 16 kHz mono.
- Listener responses were checked against attention-check probes; per-listener qualification flags are included in `participant_responses.csv`.
- The `reorganized_stimuli.csv` aggregates count same/different votes per pair.

**Is the raw data saved?**
Pre-resampling source audio is not included in the release; the IRB-covered raw participant responses are included as `participant_responses.csv`.

## Uses

**Has the dataset been used for any tasks already?**
Yes. The accompanying NeurIPS 2026 paper benchmarks 10 publicly available speaker and speech-representation models against the four evaluation tasks (continuous `P(same)` prediction, binary verification, RSA, real-to-synthetic transfer).

**Is there a repository linking to papers using the dataset?**
Will be maintained at the public repository to be created at camera-ready.

**What other tasks could the dataset be used for?**
- Listener-conditioned identity modeling.
- Voice-clone perceptual evaluation for TTS development (under the non-commercial license).
- Calibration and uncertainty estimation in speaker verification.
- Cross-language perception research, using these stimuli as a comparison anchor.

**Is there anything that should not be done with the dataset?**
- Do not use the audio, judgments, or derived embeddings for commercial purposes (CC-BY-NC 4.0).
- Do not extract clone-generation recipes from the released audio. The benchmark intentionally does not release voice-conversion or TTS training targets.
- Do not deanonymize listeners.

## Distribution

**Will the dataset be distributed to third parties?**
Yes, openly via Hugging Face under CC-BY-NC 4.0.

**How will it be distributed?**
Hugging Face dataset (anonymized URL during review, permanent URL at camera-ready). Croissant metadata accompanies the release.

**When will it be distributed?**
Initial public release: 2026-05-06 (NeurIPS submission).

**Will the dataset be distributed under a copyright or other intellectual property license?**
- Audio, judgments, and embeddings: CC-BY-NC 4.0.
- Code: MIT.
- Pretrained model weights retain their original licenses.

**Have any third parties imposed restrictions on the data?**
No.

## Maintenance

**Who will maintain the dataset?**
The authors. Contact information will be filled in at camera-ready.

**How can the maintainers be contacted?**
A mailing address will be provided at camera-ready. During review, author identification is withheld.

**Is there an erratum?**
Not yet. Errata will be tracked in `CHANGELOG.md`.

**Will the dataset be updated?**
Yes. Updates will be versioned. Older versions will remain accessible.

**Will older versions of the dataset continue to be supported?**
Yes, by maintaining tagged releases on the public repository.

**If others want to extend the dataset, is there a mechanism for them to do so?**
Yes. Researchers can collect additional listener judgments using the protocol documented in `docs/annotation_protocol.md` and contribute via the public repository (instructions to be added at camera-ready).