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.

Author names and permanent links will be added with the camera-ready version of the paper.

---

What changed in version 1.1

Version 1.1 (2026-09-30) adds the material we committed to during the NeurIPS 2026 review. The full list is in CHANGELOG.md.

• CosyVoice 2 extension subset. We added the held-out validation set in data/experiment2/. It has 10 new speakers, clones from a second TTS system (CosyVoice 2), 110 audio clips, and 1,504 judgments from 152 new listeners.
• WeSpeaker ResNet34-LM embeddings. We added an eleventh embedding model for the main set, the samples, and the extension subset.
• Hearing ability. participant_responses.csv has a new last column, hearing, with each listener's self-rated hearing on a five-point scale.
• Corrected vote counts. The vote counts in stimuli.csv were out of date for 492 of the 9,800 pairs. We recounted them from participant_responses.csv, which is the table the paper's analyses use, so no reported number changes.
• Documentation. We documented the exact clone and morph generation settings, the morph-scale scheme, the listening-study procedure, and a repository map.

Version 1.0, the version the reviewers examined, stays available under the git tag v1.0.

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	11 models (the 10 in the paper plus WeSpeaker ResNet34-LM)
Per-layer embeddings	5 models (wav2vec 2.0, HuBERT, WavLM, XLS-R, Whisper)
CosyVoice 2 extension subset	10 new speakers, 110 clips, 60 analysed pairs, 1,504 judgments, 152 listeners

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 vote counts, stimulus type, morph scale
  participant_responses.csv      # 124,876 rows: per-judgment records with listener demographics
  stimuli_interpol.csv           # 8,100 rows: exact morph weights and transcripts for Type 6
  audio/
    reference/                   # 100 *R.wav
    comparison/                  # 9,800 *.wav
  embeddings/
    rawnet3.npz, ecapa_tdnn.npz, titanet.npz, xvector.npz, resemblyzer.npz,
    wespeaker_resnet34_lm.npz,
    wav2vec2.npz, hubert.npz, wavlm.npz, xlsr.npz, whisper.npz
    layers/                      # per-layer (mean-pooled) embeddings
      wav2vec2.npz, hubert.npz, wavlm.npz, xlsr.npz, whisper.npz
  experiment2/                   # CosyVoice 2 extension subset (see below)
    speakers.csv, pairs.csv, clips.csv, responses.csv
    audio/reference/, audio/comparison/
    embeddings/*.npz, embeddings/layers/*.npz
samples/                         # 5-speaker quick-look subset (~290 MB)
code/                            # extraction scripts, analysis notebook, reproduce.sh
docs/                            # annotation protocol, schemas, model table, reproduction,
                                 # extension subset

For column-level dictionaries see docs/data_dictionary.md. For the six stimulus types and the generation settings see docs/stimulus_types.md. For the listening-study protocol see docs/annotation_protocol.md. For the extension subset see docs/extension_cosyvoice2.md.

All audio is mono 16-bit PCM WAV at 16 kHz, except data/audio/reference/F04R.wav, which is mono 16-bit PCM at 44.1 kHz. The extraction scripts resample that file to 16 kHz in memory.

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 11 main embeddings and (num_layers, embedding_dim) for the per-layer bundles. The 9,900 keys cover 100 references plus 9,800 comparisons.

One WeSpeaker embedding is NaN. The clip 6_M07B_M08B_028 is 0.09 seconds long, which is shorter than the model's minimum input, so that one pair has no WeSpeaker score.

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):

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

Repository map

The table lists the file behind each analysis input and the analyses that use it.

Analysis input	File	Analyses that use it
Speaker IDs, names, sociophonetic group, gender, age bracket	data/speakers.csv	Subgroup analyses, cross-validation folds, RSA (Task 3)
Pair list, stimulus type, metadata label, morph scale, vote counts	data/stimuli.csv	All four tasks, per-type analyses
Exact morph weights and morph transcripts	data/stimuli_interpol.csv	Morph-scale analyses
Individual judgments, familiarity flag, listener demographics and hearing	data/participant_responses.csv	P(same) target, noise ceiling, per-type reliability, familiarity analyses, human baseline
Utterance-level embeddings (11 models)	data/embeddings/{model}.npz	All four tasks
Per-layer SSL and Whisper embeddings	data/embeddings/layers/{model}.npz	Nested best-layer selection, layer analysis
Audio	data/audio/reference/, data/audio/comparison/	Re-extracting embeddings
Cross-validation folds	Built in code/benchmark_analysis.ipynb from speakers.csv. The code shuffles the 50 male and 50 female IDs with seed 42 and deals them into 10 folds of 5 men and 5 women.	All cross-validated analyses
Analysis for the paper's tables and figures	code/benchmark_analysis.ipynb, run by code/reproduce.sh	All four tasks
Embedding extraction	code/extract_*.py, code/extract_ssl_layers.py, code/run_all_extractions.sh	Rebuilding data/embeddings/
CosyVoice 2 extension subset	data/experiment2/, code/extension_cosyvoice2.py	Replication with a second TTS system
Five-speaker sample	samples/	Quick inspection

CosyVoice 2 extension subset

The extension subset in data/experiment2/ is a held-out validation set. We use it to test whether the benchmark's findings hold for clones from a different TTS system. It has 10 public-figure speakers who are not among the 100 main speakers, and each one is paired with a soundalike, a different public figure whose voice is widely considered similar. The folder is named experiment2 because the set was collected in a second listening experiment.

• Audio. 110 clips, i.e., 10 reference clips, 60 comparison clips that listeners judged, and 40 extra genuine clips of the 10 speakers that no listener judged.
• Pairs. Each speaker has six judged pairs, i.e., a different recording of the same speaker, two CosyVoice 2 clones of the speaker, a real recording of the soundalike, and two CosyVoice 2 clones of the soundalike. pairs.csv also lists 10 same-recording records whose comparison audio is identical to the reference clip, and these records are not part of the analysed set.
• Judgments. 1,504 ratings from 152 listeners recruited on Prolific, on a four-point scale from 1 (definitely different) to 4 (definitely the same). The 60 analysed pairs have 1,291 ratings, with a median of 21 ratings per pair. We count ratings of 3 and 4 as "same" when we compute P(same).
• Embeddings. The same 11 models and 5 per-layer bundles as the main set, keyed by clip_id.

code/extension_cosyvoice2.py computes P(same), scores all models with frozen settings, and reports each model's Pearson r with a speaker bootstrap interval and the rank correlation with the main benchmark. docs/extension_cosyvoice2.md documents the files, the generation settings, and the listening study.

Quick start

pip install -r requirements.txt
cd code && bash reproduce.sh          # ~10 min from cached embeddings
python3 extension_cosyvoice2.py       # CosyVoice 2 extension subset, ~30 seconds

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). The clips are conversational speech. The mean clip length is 11.6 seconds (13.1 seconds for references and 11.6 seconds for comparisons).
• Voice clones. Generated with Cartesia Sonic 3, 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 voice of the reference speaker is mixed with the voice of each of 4 comparison speakers from the same demographic cell (sociophonetic group, age group, and gender), at 10 mixing weights for each of 2 source recordings, plus 1 anchor at weight 1. This yields 4 x 2 x 10 + 1 = 81 Type 6 stimuli per reference speaker (8,100 total). The morphs were made with the voice-mixing feature of the same Cartesia system. docs/stimulus_types.md gives the exact weights and settings.
• 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. Before the task, listeners reported their gender, age bracket, first language, and hearing ability.

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. The listener fields are limited to the study answers and the self-reported gender, age bracket, first language, and hearing ability. Listener identifiers in the extension subset are sequential labels (L001 to L152), and no Prolific identifier is included.

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.
• All clones and morphs in the main set come from one commercial system (Cartesia Sonic 3). The extension subset adds clones from a second system (CosyVoice 2), but it has only 10 speakers.

Responsible use

We document the settings we used to generate the clones and morphs, so that others can check how the stimuli were made. We don't release methods for making clones that fool listeners or verification systems more often, and we don't release 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 with the camera-ready version.

@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   = {Author names to be added with the camera-ready version.}
}

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.