{ "@context": { "@language": "en", "@vocab": "https://schema.org/", "citeAs": "cr:citeAs", "column": "cr:column", "conformsTo": "dct:conformsTo", "cr": "http://mlcommons.org/croissant/", "rai": "http://mlcommons.org/croissant/RAI/", "prov": "http://www.w3.org/ns/prov#", "data": { "@id": "cr:data", "@type": "@json" }, "dataType": { "@id": "cr:dataType", "@type": "@vocab" }, "dct": "http://purl.org/dc/terms/", "equivalentProperty": "cr:equivalentProperty", "examples": { "@id": "cr:examples", "@type": "@json" }, "extract": "cr:extract", "field": "cr:field", "fileProperty": "cr:fileProperty", "fileObject": "cr:fileObject", "fileSet": "cr:fileSet", "format": "cr:format", "includes": "cr:includes", "isLiveDataset": "cr:isLiveDataset", "jsonPath": "cr:jsonPath", "key": "cr:key", "md5": "cr:md5", "parentField": "cr:parentField", "path": "cr:path", "recordSet": "cr:recordSet", "references": "cr:references", "regex": "cr:regex", "repeated": "cr:repeated", "replace": "cr:replace", "samplingRate": "cr:samplingRate", "sc": "https://schema.org/", "separator": "cr:separator", "source": "cr:source", "subField": "cr:subField", "transform": "cr:transform" }, "@type": "sc:Dataset", "name": "VIPBench", "description": "A human-aligned benchmark for voice identity perception. 124,876 same/different identity judgments from 1,290 English-speaking listeners on 9,800 voice pairs across 100 demographically-stratified speakers, spanning real recordings, AI voice clones, and continuously morphed voices. Includes pre-extracted embeddings for 10 speaker and speech-representation models.", "conformsTo": "http://mlcommons.org/croissant/1.0", "citeAs": "Anonymous Authors. VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning. Advances in Neural Information Processing Systems Datasets and Benchmarks Track, 2026. Anonymized for double-blind review.", "license": "https://creativecommons.org/licenses/by-nc/4.0/", "url": "https://huggingface.co/datasets/anonymous/vipbench", "version": "1.0.0", "datePublished": "2026-05-06", "keywords": [ "speaker embeddings", "voice identity perception", "human-aligned benchmark", "voice cloning", "perceptual evaluation", "speaker verification", "speech representations" ], "creator": { "@type": "Organization", "name": "Anonymous (NeurIPS 2026 double-blind review)" }, "isLiveDataset": false, "distribution": [ { "@type": "cr:FileObject", "@id": "vipbench-bundle", "name": "vipbench-bundle", "description": "VIPBench dataset bundle on Hugging Face. The bundle is the dataset git repository; FileObjects and FileSets below are paths within it.", "contentUrl": "https://huggingface.co/datasets/sendfuze/vipbench", "encodingFormat": "git+https", "sha256": "TO_BE_COMPUTED_BY_PUBLISHER" }, { "@type": "cr:FileObject", "@id": "speakers-csv", "name": "speakers.csv", "description": "Per-speaker metadata for the 100 celebrity speakers.", "contentUrl": "data/speakers.csv", "encodingFormat": "text/csv", "sha256": "TO_BE_COMPUTED_BY_PUBLISHER", "containedIn": { "@id": "vipbench-bundle" } }, { "@type": "cr:FileObject", "@id": "stimuli-csv", "name": "stimuli.csv", "description": "Per-pair aggregates for the 9,800 voice pairs.", "contentUrl": "data/stimuli.csv", "encodingFormat": "text/csv", "sha256": "TO_BE_COMPUTED_BY_PUBLISHER", "containedIn": { "@id": "vipbench-bundle" } }, { "@type": "cr:FileObject", "@id": "responses-csv", "name": "participant_responses.csv", "description": "Per-judgment records (124,876 rows).", "contentUrl": "data/participant_responses.csv", "encodingFormat": "text/csv", "sha256": "TO_BE_COMPUTED_BY_PUBLISHER", "containedIn": { "@id": "vipbench-bundle" } }, { "@type": "cr:FileObject", "@id": "stimuli-interpol-csv", "name": "stimuli_interpol.csv", "description": "Type 6 morph trajectory metadata (8,100 rows).", "contentUrl": "data/stimuli_interpol.csv", "encodingFormat": "text/csv", "sha256": "TO_BE_COMPUTED_BY_PUBLISHER", "containedIn": { "@id": "vipbench-bundle" } }, { "@type": "cr:FileSet", "@id": "audio-reference", "name": "audio/reference", "description": "100 reference audio clips (16 kHz mono WAV), one per speaker.", "containedIn": { "@id": "vipbench-bundle" }, "encodingFormat": "audio/wav", "includes": "data/audio/reference/*.wav" }, { "@type": "cr:FileSet", "@id": "audio-comparison", "name": "audio/comparison", "description": "9,800 comparison audio clips (16 kHz mono WAV).", "containedIn": { "@id": "vipbench-bundle" }, "encodingFormat": "audio/wav", "includes": "data/audio/comparison/*.wav" }, { "@type": "cr:FileSet", "@id": "embeddings-main", "name": "embeddings", "description": "Pre-extracted utterance-level embeddings for 10 speaker and speech-representation models. Each .npz contains 9,900 keys (audio basenames) mapped to 1-D embedding vectors.", "containedIn": { "@id": "vipbench-bundle" }, "encodingFormat": "application/x-npz", "includes": "data/embeddings/*.npz" }, { "@type": "cr:FileSet", "@id": "embeddings-layers", "name": "embeddings/layers", "description": "Per-transformer-layer mean-pooled embeddings for 5 SSL models (wav2vec 2.0, HuBERT, WavLM, XLS-R, Whisper). Each .npz contains 9,900 keys mapped to 2-D arrays of shape (num_layers, embedding_dim).", "containedIn": { "@id": "vipbench-bundle" }, "encodingFormat": "application/x-npz", "includes": "data/embeddings/layers/*.npz" } ], "recordSet": [ { "@type": "cr:RecordSet", "@id": "speakers", "name": "speakers", "description": "100 celebrity speakers stratified across 5 sociophonetic groups x 2 genders x 2 age brackets.", "field": [ { "@type": "cr:Field", "@id": "speakers/id", "name": "id", "description": "Speaker identifier (e.g., F01, M07).", "dataType": "sc:Text", "source": { "fileObject": { "@id": "speakers-csv" }, "extract": { "column": "id" } } }, { "@type": "cr:Field", "@id": "speakers/name", "name": "name", "description": "Speaker name (public figure).", "dataType": "sc:Text", "source": { "fileObject": { "@id": "speakers-csv" }, "extract": { "column": "name" } } }, { "@type": "cr:Field", "@id": "speakers/group", "name": "group", "description": "Sociophonetic group (1=New York City English, 2=Southern American English, 3=African American English, 4=Latino English, 5=Asian American English).", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "speakers-csv" }, "extract": { "column": "group" } } }, { "@type": "cr:Field", "@id": "speakers/gender", "name": "gender", "description": "Speaker gender (1=male, 2=female).", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "speakers-csv" }, "extract": { "column": "gender" } } }, { "@type": "cr:Field", "@id": "speakers/age", "name": "age", "description": "Speaker age bracket (1=under 45, 2=55 or older).", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "speakers-csv" }, "extract": { "column": "age" } } } ] }, { "@type": "cr:RecordSet", "@id": "stimuli", "name": "stimuli", "description": "9,800 voice pairs with per-pair human-judgment aggregates and metadata.", "field": [ { "@type": "cr:Field", "@id": "stimuli/id", "name": "id", "description": "Stimulus identifier (matches the audio basename for the comparison clip).", "dataType": "sc:Text", "source": { "fileObject": { "@id": "stimuli-csv" }, "extract": { "column": "id" } } }, { "@type": "cr:Field", "@id": "stimuli/stimuli_type", "name": "stimuli_type", "description": "Stimulus type (1-6). See docs/stimulus_types.md.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "stimuli-csv" }, "extract": { "column": "stimuli_type" } } }, { "@type": "cr:Field", "@id": "stimuli/reference", "name": "reference", "description": "Reference speaker ID (joins to speakers/id).", "dataType": "sc:Text", "source": { "fileObject": { "@id": "stimuli-csv" }, "extract": { "column": "reference" } }, "references": { "field": { "@id": "speakers/id" } } }, { "@type": "cr:Field", "@id": "stimuli/comparison", "name": "comparison", "description": "Comparison speaker ID for non-Type-6 pairs.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "stimuli-csv" }, "extract": { "column": "comparison" } } }, { "@type": "cr:Field", "@id": "stimuli/voice_clone", "name": "voice_clone", "description": "Whether the comparison clip is an AI voice clone (1) or natural recording (0).", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "stimuli-csv" }, "extract": { "column": "voice_clone" } } }, { "@type": "cr:Field", "@id": "stimuli/correct_answer", "name": "correct_answer", "description": "Metadata-label same/different (1=same speaker by metadata, 0=different).", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "stimuli-csv" }, "extract": { "column": "correct_answer" } } }, { "@type": "cr:Field", "@id": "stimuli/scale", "name": "scale", "description": "For Type 6 morphs, the interpolation scale (0=reference voice, 100=other voice). 100 for non-morph pairs.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "stimuli-csv" }, "extract": { "column": "scale" } } }, { "@type": "cr:Field", "@id": "stimuli/num_response", "name": "num_response", "description": "Number of listener judgments collected for this pair.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "stimuli-csv" }, "extract": { "column": "num_response" } } }, { "@type": "cr:Field", "@id": "stimuli/same_vote", "name": "same_vote", "description": "Number of listeners who judged the pair as the same speaker.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "stimuli-csv" }, "extract": { "column": "same_vote" } } }, { "@type": "cr:Field", "@id": "stimuli/diff_vote", "name": "diff_vote", "description": "Number of listeners who judged the pair as different speakers.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "stimuli-csv" }, "extract": { "column": "diff_vote" } } } ] }, { "@type": "cr:RecordSet", "@id": "responses", "name": "participant_responses", "description": "124,876 individual listener judgments.", "field": [ { "@type": "cr:Field", "@id": "responses/user_id", "name": "user_id", "description": "Pseudonymized listener identifier.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "responses-csv" }, "extract": { "column": "user_id" } } }, { "@type": "cr:Field", "@id": "responses/stimuli_id", "name": "stimuli_id", "description": "Stimulus identifier (joins to stimuli/id).", "dataType": "sc:Text", "source": { "fileObject": { "@id": "responses-csv" }, "extract": { "column": "stimuli_id" } }, "references": { "field": { "@id": "stimuli/id" } } }, { "@type": "cr:Field", "@id": "responses/stimuli_type", "name": "stimuli_type", "description": "Stimulus type (1-6).", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "responses-csv" }, "extract": { "column": "stimuli_type" } } }, { "@type": "cr:Field", "@id": "responses/answer", "name": "answer", "description": "Listener's binary same/different judgment (1=same, 0=different).", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "responses-csv" }, "extract": { "column": "answer" } } }, { "@type": "cr:Field", "@id": "responses/correct", "name": "correct", "description": "Whether the listener's answer matches the metadata label.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "responses-csv" }, "extract": { "column": "correct" } } }, { "@type": "cr:Field", "@id": "responses/know_speaker", "name": "know_speaker", "description": "Listener-recognition flag (1=listener identified the reference speaker).", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "responses-csv" }, "extract": { "column": "know_speaker" } } } ] } ], "rai:dataCollection": "Listener judgments were collected via the Centaur AI online crowdsourcing platform under an Institutional Review Board (IRB) approved research protocol. Each pair received at least 8 judgments from English-speaking adult participants. Source audio was selected from publicly available recordings of 100 US celebrity speakers; AI voice clones were generated from these source clips using a state-of-the-art text-to-speech system; voice morphs were generated by interpolating a voice-conversion latent between two source speakers.", "rai:dataCollectionType": [ "Crowdsourcing", "Synthetic data generation" ], "rai:hasSyntheticData": true, "prov:wasDerivedFrom": [ { "@id": "https://huggingface.co/datasets/sendfuze/vipbench", "prov:label": "VIPBench per-speaker source clips (R + A-E variants)", "description": "For each of 100 US celebrity speakers: one reference clip (R) plus 5 additional source clips (variants A through E). All clips are excerpts of publicly available interview and podcast recordings, curated by the dataset authors under an IRB-approved research protocol. Reference (R) clips are released in data/audio/reference/. Voice clones (stimulus types 3 and 5) are seeded from one of the speaker's A through E source clips; voice morphs (stimulus type 6) interpolate between A through E source clips of two within-group speakers. No aggregated external dataset URI is claimed. Attribution: Anonymous (NeurIPS 2026 double-blind review). Dataset authors; identity withheld for review.", "license": "https://creativecommons.org/licenses/by-nc/4.0/" }, { "@id": "https://www.cartesia.ai", "prov:label": "Cartesia text-to-speech system", "description": "State-of-the-art commercial TTS system used to synthesize the voice clones (stimulus types 3 and 5) and voice morphs (stimulus type 6). Seed inputs to the synthesis system are the per-speaker source clips described above. Attribution: Cartesia." } ], "prov:wasGeneratedBy": [ { "@type": "prov:Activity", "prov:label": "Source-clip collection", "prov:atTime": "2025-2026", "description": "Reference (R) and 5 additional source clips (variants A through E) per speaker were curated from publicly available interview and podcast recordings of 100 US celebrity speakers, stratified across 5 sociophonetic groups (1=NYC English, 2=Southern American English, 3=African American English, 4=Latino English, 5=Asian American English) x 2 genders x 2 age brackets (under 45, 55 or older), 5 speakers per cell. Conducted under IRB approval. Attribution: Anonymous (NeurIPS 2026 double-blind review) (Dataset authors; identity withheld for review.)." }, { "@type": "prov:Activity", "prov:label": "Voice-clone synthesis", "prov:atTime": "2025-2026", "description": "Voice clones (stimulus types 3 and 5) generated using Cartesia, a state-of-the-art commercial TTS system, seeded from a natural source clip of the speaker being cloned. Each Type 3 clone shares its seed clip with the comparison clip of the matched Type 2 pair; each Type 5 clone shares its seed clip with the matched Type 4 pair (e.g., 3_F01B is seeded from the same F01B source clip used as the comparison in 2_F01B). Attribution: Cartesia." }, { "@type": "prov:Activity", "prov:label": "Voice-morph synthesis", "prov:atTime": "2025-2026", "description": "Voice morphs (stimulus type 6) generated using the voice-morphing feature of the same Cartesia system. For each of the 100 reference speakers, the latent voice representation 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, yielding 4 x 2 x 10 + 1 = 81 stimuli per reference speaker (8,100 total). Attribution: Cartesia." }, { "@type": "prov:Activity", "prov:label": "Audio preprocessing", "prov:atTime": "2025-2026", "description": "All audio (reference, comparison, clones, morphs) resampled to 16 kHz mono float32 WAV. Clips trimmed to approximately 6 seconds. Attribution: Anonymous (NeurIPS 2026 double-blind review)." }, { "@type": "prov:Activity", "prov:label": "Human annotation (listening study)", "prov:atTime": "2025-2026", "description": "1,290 English-speaking adult crowdworkers recruited via the Centaur AI platform answered binary same/different identity judgments on 9,800 voice pairs. Each pair received at least 8 judgments (median 10, range 8 to 92). Each trial presented reference + 1-second silence + short beep + comparison; listeners answered (a) binary same/different and (b) an optional speaker-recognition probe. Conducted under IRB approval with informed consent. Attribution: Centaur AI crowdsourcing platform; 1,290 adult English-speaking crowdworkers (Listener identifiers in data/participant_responses.csv are pseudonymized integers tied to no external account.)." }, { "@type": "prov:Activity", "prov:label": "Aggregation and reliability analysis", "prov:atTime": "2025-2026", "description": "Per-pair P(same) and same/different vote counts computed by aggregating individual responses (data/stimuli.csv). Spearman-Brown corrected split-half reliability rho_SB = 0.705 computed over 100 random splits of the 1,290-participant pool. Per-listener attention-check qualification flags shipped alongside raw responses (data/participant_responses.csv). Attribution: Anonymous (NeurIPS 2026 double-blind review)." }, { "@type": "prov:Activity", "prov:label": "Embedding extraction", "prov:atTime": "2025-2026", "description": "Utterance-level embeddings extracted from each of 9,900 audio files using 10 publicly available pretrained models (x-vector, ECAPA-TDNN, RawNet3, TitaNet, resemblyzer, wav2vec 2.0, HuBERT, WavLM, XLS-R, Whisper). For self-supervised models, both final-layer and per-transformer-layer mean-pooled embeddings are released. Extraction scripts are in code/extract_*.py. Attribution: Anonymous (NeurIPS 2026 double-blind review)." } ], "rai:dataCollectionRawData": "Per-speaker source clips: for each of 100 US celebrity speakers, one reference clip (data/audio/reference/R.wav) plus five additional clips labeled A through E (e.g., F01A, F01B, ..., F01E). All source clips are excerpts of publicly available interview and podcast recordings, curated directly by the dataset authors under an IRB-approved research protocol and resampled to 16 kHz mono float32 WAV. The 100 reference (R) clips are released in data/audio/reference/. The A through E source clips for the speakers being cloned or morphed are not redistributed as standalone files in this release; their identifiers appear in stimulus IDs (e.g., '3_F01B' indicates a Type 3 clone of speaker F01 generated from clip F01B). Speaker metadata is in data/speakers.csv. No aggregated external dataset URI is claimed. | Voice clones (stimulus types 3 and 5; data/audio/comparison/3_*.wav and 5_*.wav): synthetically generated using the Cartesia text-to-speech system (https://www.cartesia.ai), 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., '3_F01B' is seeded from the same F01B source clip used as the comparison in '2_F01B'). Type 3 = same-speaker clone (cloned voice paired with the reference of the same speaker). Type 5 = different-speaker clone (cloned voice paired with the reference of a different speaker). | Voice morphs (stimulus type 6; data/audio/comparison/6_*.wav): synthetically generated using the voice-morphing feature of the same Cartesia system. 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, with 10 morph scales between 0 and 1 plus 1 anchor at scale 1, yielding 4 x 2 x 10 + 1 = 81 stimuli per reference speaker (8,100 total). Stimulus IDs encode the two endpoint speakers, the seed-recording variants, and the scale (e.g., '6_M05A_M03A_065' = morph between M05 and M03 with seed recordings A from each, at scale 65). Per-stimulus trajectory metadata (endpoint speakers, seed-recording variants, scale) is in data/stimuli_interpol.csv. | Pretrained model checkpoints used to compute the embeddings in data/embeddings/ are listed in docs/model_table.md and retain their original licenses; see https://huggingface.co/datasets/sendfuze/vipbench/blob/main/docs/model_table.md.", "rai:provenanceActivities": "[] Per-speaker source clips for 100 US celebrity speakers collected from publicly available interview and podcast recordings: one reference clip (R) plus five additional source clips (A through E) per speaker. Speakers are stratified across 5 sociophonetic groups x 2 genders x 2 age brackets (5 speakers per cell). Curated by the dataset authors under an IRB-approved research protocol. | [] Voice clones (stimulus types 3 and 5) generated by Cartesia TTS, seeded from a natural source clip of the speaker being cloned. The seed clip for a Type 3 clone is the same source clip used as the comparison in the matched Type 2 pair; the seed clip for a Type 5 clone is the same source clip used as the comparison in the matched Type 4 pair. Voice morphs (stimulus type 6) generated using the voice-morphing feature of the same Cartesia system: for each of 100 reference speakers, the latent voice representation 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, yielding 4 x 2 x 10 + 1 = 81 stimuli per reference speaker (8,100 total). | [] All audio resampled to 16 kHz mono float32 WAV. Clips trimmed to approximately 6 seconds. Speaker IDs assigned in the form F01-F50, M01-M50. | [] 1,290 English-speaking adult crowdworkers recruited via the Centaur AI platform answered binary same/different identity judgments on 9,800 voice pairs (median 10 judgments per pair, range 8 to 92). Each trial presented reference + 1-second silence + beep + comparison; listeners answered (a) binary same/different and (b) optional speaker recognition. Conducted under IRB approval with informed consent. | [] Per-pair P(same) and same/different vote counts computed by aggregating individual responses (data/stimuli.csv). Per-listener attention-check qualification flags computed and shipped alongside raw responses (data/participant_responses.csv). Spearman-Brown corrected split-half reliability rho_SB = 0.705 computed over 100 random splits of the 1,290-participant pool. | [] Utterance-level embeddings extracted from each of 9,900 audio files using 10 publicly available pretrained models (x-vector, ECAPA-TDNN, RawNet3, TitaNet, resemblyzer, wav2vec 2.0, HuBERT, WavLM, XLS-R, Whisper). For self-supervised models, both the final-layer embedding and per-transformer-layer mean-pooled embeddings are released. Extraction scripts are in code/extract_*.py.", "rai:dataCollectionMissingValues": "The know_speaker field is missing for some early-trial responses (less than 1% of records). Listeners with fewer than the qualification threshold of attention-check passes are flagged but their responses are still released.", "rai:dataCollectionTimeFrame": { "@type": "DateTime", "@value": "2025-01-01/2026-04-30" }, "rai:dataAnnotationProtocol": "Each trial presents a single audio clip in which a reference recording is followed by 1 second of silence, a short beep, and a comparison recording. The listener answers (a) whether the two clips came from the same speaker (binary same/different) and (b) optionally identifies which of four within-group celebrities (or 'I don't know') they recognize in the reference clip. Stimulus presentation order is randomized within participant. Participants provide informed consent prior to participating; no deception is involved.", "rai:dataAnnotationPlatform": "Centaur AI crowdsourcing platform (https://centaur.ai)", "rai:dataAnnotationAnalysis": "Per-pair P(same) is computed as the fraction of listeners who judged the pair as the same speaker. Inter-rater agreement is reported as Spearman-Brown corrected split-half reliability (rho_SB = 0.705 over 100 random splits).", "rai:annotationsPerItem": "median 10, range 8 to 92", "rai:dataPreprocessingProtocol": "Audio was resampled to 16 kHz mono. Listener responses were validated against attention-check probes embedded in the study. The reorganized_stimuli.csv aggregates count same/different votes per pair.", "rai:dataReleaseMaintenancePlan": "The dataset is versioned (v1.0 at NeurIPS 2026 submission). Updates are tracked in CHANGELOG.md and tagged on the public repository. The authors will respond to errata via the public repository issue tracker (link to be added at camera-ready).", "rai:dataLifecycleStage": "Released", "rai:personalSensitiveInformation": "The dataset names public-figure speakers because the celebrity-stratified design is integral to the benchmark and source recordings are publicly available. Listener identifiers are pseudonymized integers tied to no external account. Listener demographic fields are limited to age band, gender, and a binary first-language flag.", "rai:dataSocialImpact": "Voice-cloning systems that better align with human perception could inform adversarial use such as more convincing fraudulent calls. The same alignment knowledge also strengthens defenses: perception-aligned identity models can flag voice clones that metadata-based verification would accept. The benchmark measures model-human alignment at the evaluation level and does not release clone-generation recipes or adversarial training targets.", "rai:dataLimitations": [ "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." ], "rai:dataUseCases": [ "Benchmarking speaker embedding models against human voice-identity perception (Pearson r, R^2 against P(same)).", "Human-aligned binary speaker verification (AUC, Platt-calibrated ECE against the listener majority vote).", "Representational similarity analysis (RSA) between human and model representational dissimilarity matrices.", "Real-to-synthetic distribution-shift evaluation (does a predictor fit on real-speech pairs still work on voice clones and morphs?).", "Listener-conditioned identity modeling (using per-listener responses).", "Calibration and uncertainty estimation in speaker verification." ], "rai:dataBiases": "Speakers are 100 US celebrities; the dataset over-represents US-dialect, professionally-recorded speech. Listeners are English-speaking adult crowdworkers. Generalization to non-English speakers, listeners, or in-the-wild audio conditions is not measured. The 5 sociophonetic groups carry 20 speakers each, limiting subgroup statistical power." }