Tang, Xudong, Groh, Matthew · Nov 3, 2026
VIPBench measures whether speech models match human judgments of voice identity
Markdown · JSON · Original source
Authors Xudong Tang and Matthew Groh introduce VIPBench, a for comparing speech models with how listeners perceive voice identity. Unlike speaker-recognition benchmarks that test whether a model recovers a recording’s speaker label, VIPBench asks whether models align with listeners deciding if two voices sound like the same person. The benchmark includes 124,876 judgments from 1,290 listeners on 9,800 pairs involving real recordings, voice clones and morphs between speakers’ voices. In the authors’ evaluation, listeners often heard clones as different people from recordings of the same speaker: 41% of responses made that judgment. The best of eleven evaluated models explained 61% of the explainable variance, and no model reached the noise ceiling. The authors also report that models trained with speaker labels aligned better with listeners than self-supervised models, while one model with a 12% speaker-label error rate aligned as well as models below 2%. Their argument is that recovering a speaker label and matching human perception are different goals, with implications they identify for intellectual property rights, cybersecurity and everyday life. They say the benchmark enables evaluation on real, cloned and morphed speech, and their results motivate training models directly on listener judgments.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 3, 2026 (2026-11-03).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T00:04:57.742Z. First observed: 2026-11-05T05:09:29.813Z. Retrieved: 2026-11-09T13:52:30.095Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The abstract describes a using recordings from 100 public figures and judgments from 1,290 listeners; it does not report whether the findings generalize beyond them. The evidence provided is the paper’s abstract and does not give details of the listener sample or the evaluation procedures.
Open question
How well do VIPBench’s results hold for listeners and public figures beyond the 1,290 listeners and 100 public figures in this ?
Source-stated
The authors introduce VIPBench to evaluate whether speaker-representation models align with human judgments of voice identity, rather than only testing recovery of speaker labels.
We introduce the Voice Identity Perception benchmark (VIPBench)
Source-stated
VIPBench contains 124,876 judgments from 1,290 listeners on 9,800 recording pairs involving 100 public figures.
124,876 same-or-different judgments from 1,290 listeners on 9,800 recording pairs from 100 public figures
Source-stated
In 41% of responses, listeners judged a voice clone and a real recording of the same speaker to be different people.
in 41% of responses, listeners judged a clone and a real recording of the same speaker to be different people
Source-stated
Among eleven evaluated speech models, the best explained 61% of the explainable variance, and none reached the noise ceiling.
No model reaches the noise ceiling; the best model explains 61% of the explainable variance.
Source-stated
Models trained with speaker labels aligned with listener judgments better than self-supervised models, according to the authors.
Models trained with speaker labels align with listeners better than self-supervised models do
Source-stated
The authors say the results motivate training speaker models directly on listener judgments.
our results motivate training speaker models directly on listener judgments.
Source and contributor context
Tang, Xudong, Groh, Matthew · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: ac336d54-ff45-4f7d-8330-c663e4787957.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T00:04:57.742Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: eed38331fd17414ca9d28ab916fa634008c75808346405ae63966280a1b451b5. Source text hash: 2bc3056c73392e9359941be371e9d3720aabdeffa9c0c95609654efbcb92f40b.
Revision and source-check history
Version 1 · 2026-11-10T00:04:57.742Z
VIPBench measures whether speech models match human judgments of voice identity
Authors Xudong Tang and Matthew Groh introduce VIPBench, a for comparing speech models with how listeners perceive voice identity. Unlike speaker-recognition benchmarks that test whether a model recovers a recording’s speaker label, VIPBench asks whether models align with listeners deciding if two voices sound like the same person. The benchmark includes 124,876 judgments from 1,290 listeners on 9,800 pairs involving real recordings, voice clones and morphs between speakers’ voices. In the authors’ evaluation, listeners often heard clones as different people from recordings of the same speaker: 41% of responses made that judgment. The best of eleven evaluated models explained 61% of the explainable variance, and no model reached the noise ceiling. The authors also report that models trained with speaker labels aligned better with listeners than self-supervised models, while one model with a 12% speaker-label error rate aligned as well as models below 2%. Their argument is that recovering a speaker label and matching human perception are different goals, with implications they identify for intellectual property rights, cybersecurity and everyday life. They say the benchmark enables evaluation on real, cloned and morphed speech, and their results motivate training models directly on listener judgments.
Original source
cs.AI · Article
Source details
Original source date: Nov 3, 2026 (2026-11-03).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T00:04:57.742Z. First observed: 2026-11-05T05:09:29.813Z. Retrieved: 2026-11-09T13:52:30.095Z.
Limits
The abstract describes a using recordings from 100 public figures and judgments from 1,290 listeners; it does not report whether the findings generalize beyond them. The evidence provided is the paper’s abstract and does not give details of the listener sample or the evaluation procedures.
Open question
How well do VIPBench’s results hold for listeners and public figures beyond the 1,290 listeners and 100 public figures in this ?
Source-stated
The authors introduce VIPBench to evaluate whether speaker-representation models align with human judgments of voice identity, rather than only testing recovery of speaker labels.
We introduce the Voice Identity Perception benchmark (VIPBench)
Source-stated
VIPBench contains 124,876 judgments from 1,290 listeners on 9,800 recording pairs involving 100 public figures.
124,876 same-or-different judgments from 1,290 listeners on 9,800 recording pairs from 100 public figures
Source-stated
In 41% of responses, listeners judged a voice clone and a real recording of the same speaker to be different people.
in 41% of responses, listeners judged a clone and a real recording of the same speaker to be different people
Source-stated
Among eleven evaluated speech models, the best explained 61% of the explainable variance, and none reached the noise ceiling.
No model reaches the noise ceiling; the best model explains 61% of the explainable variance.
Source-stated
Models trained with speaker labels aligned with listener judgments better than self-supervised models, according to the authors.
Models trained with speaker labels align with listeners better than self-supervised models do
Source-stated
The authors say the results motivate training speaker models directly on listener judgments.
our results motivate training speaker models directly on listener judgments.
Source and contributor context
Tang, Xudong, Groh, Matthew · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: ac336d54-ff45-4f7d-8330-c663e4787957.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T00:04:57.742Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: eed38331fd17414ca9d28ab916fa634008c75808346405ae63966280a1b451b5. Source text hash: 2bc3056c73392e9359941be371e9d3720aabdeffa9c0c95609654efbcb92f40b.
passed · 2026-11-10T00:04:57.742Z · anthropic/claude-opus-5.5
Publication: ac336d54-ff45-4f7d-8330-c663e4787957
attribution — pass: The draft correctly attributes the paper to Xudong Tang and Matthew Groh. It frames the findings and the training recommendation as the authors' reported results and argument, not as established fact.
Authors:Xudong Tang, Matthew Groh
summary — pass: The draft first explains the development: a benchmark built on human same-or-different judgments, in contrast to speaker-label tests. It then gives the key results and the authors' stated stakes and recommendation. Terms such as noise ceiling, explainable variance and self-supervised are kept exactly and are suited to tooltips. The core meaning stays readable without them. The question raises a real generalization gap the abstract does not resolve.
However, these benchmarks do not evaluate how humans perceive voice identity.
date — pass: The originalPublished date of 2026-11-03 matches the v1 arXiv submission date. It is not the later feed date of 2026-11-05.
[Submitted on 3 Nov 2026]
support — pass: Every numeric claim matches the abstract exactly, as do the model comparisons and the stated motivation. These include 124,876 judgments, 1,290 listeners, 9,800 pairs, 41%, eleven models, 61%, and error rates of 12% versus below 2%. The implications for IP rights, cybersecurity and everyday life come from the source itself. The draft does not invent beneficiaries or costs.
a model with a 12% error rate on speaker labels aligns with listeners as well as models with error rates below 2%
media — pass: The selected media is the verified arXiv abstract page for this paper. It is the direct original source, and no better source media was found.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference