---
id: MR-0003
type: metareport
slug: voice-cloning-human-detection
title: 'Voice Clones and Synthetic Speech: How Well Do People Detect Them?'
question: Can people tell cloned or synthetic voices from real speakers, and what changes that ability?
version: 0.1.0
published: 2026-10-19
updated: 2026-10-19
status: seed-partially-verified
method: narrative-synthesis
scope: Peer-reviewed behavioural studies (2023-2026) of human listeners judging synthetic or cloned speech against real speech, plus the audio subset of one meta-analysis. Split out of MR-0001 on 2026-10-19; MR-0001.F12 and F15 remain the cross-modal summary.
sources:
- barrington-2025-voice-clones
- lavan-2025-voice-clones-hyperrealism
- ivory-2026-faces-voices-domain
- mai-2023-speech-deepfakes
- san-segundo-2026-audio-language-style
- groh-2024-political-speech
- diel-2024-deepfake-detection-meta
- cooke-2024-coin-toss
related:
- MR-0001
- MR-0002
tags:
- voice-cloning
- synthetic-speech
- vishing
- human-detection
- audio-deepfakes
verification_counts:
  verified-fulltext: 5
  verified-abstract: 3
  verified-doi: 0
  unverified: 0
canonical_url: https://research.breacher.ai/metareports/voice-cloning-human-detection.md
---

# MR-0003 · Voice Clones and Synthetic Speech: How Well Do People Detect Them?

> **Question:** Can people tell cloned or synthetic voices from real speakers, and what changes that ability?  
> **Answer in one line:** Barely, for modern clones of unfamiliar speakers: a commercial clone was accepted as the same person in a median 83.3% of pairs, listener sensitivity was d′ ≈ 0.1–0.2, and confidence did not track accuracy. Older or generic synthetic voices are easier (≈70% single-clip), and comparing against genuine reference audio helps most.  
> **Version:** 0.1.0 · **Updated:** 2026-10-19 · **Sources:** 8 (5 checked against full text, 3 against the official abstract, 0 DOI-only, 0 unverified; see [Verification status](#verification)).

## TL;DR {#tldr}

- **Clones pass as the person:** real–clone pairs judged the same speaker in a median 83.3% (IQR 69.2–100.0); real–real pairs 100.0% [barrington-2025-voice-clones](/data/citations.json#barrington-2025-voice-clones).
- **Near-chance sensitivity:** d′ 0.10–0.13 for ElevenLabs clones [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism); 53.95% correct (d′ .20) [ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain); clones identified 60.8% [barrington-2025-voice-clones](/data/citations.json#barrington-2025-voice-clones).
- **Bias toward "human":** criterion c = −0.65 for clones [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism); c = −.07 [ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain).
- **Realism parity, not hyperrealism:** clones rated as real as human voices (63.0 vs 63.8), not more [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism).
- **Older and generic voices are easier:** VITS TTS 70.35% single clip [mai-2023-speech-deepfakes](/data/citations.json#mai-2023-speech-deepfakes); generic AI voices d′ 2.93 [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism); TTS fakes caught 72% vs 83% for voice actors [groh-2024-political-speech](/data/citations.json#groh-2024-political-speech).
- **Comparison helps:** picking the fake from a pair 85.59% vs 70.35% single clip; familiarisation only +3.84 points [mai-2023-speech-deepfakes](/data/citations.json#mai-2023-speech-deepfakes).
- **Confidence is no guide:** metacognitive AUROC .48 for voices [ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain).

## Key numbers {#key-numbers}

| Metric | Value | Source | Verification |
|---|---|---|---|
| Real–clone pairs judged same speaker (median) | 83.3% (IQR 69.2–100.0) | barrington-2025-voice-clones | full text |
| Clones / real clips correctly classified | 60.8% / 67.4% | barrington-2025-voice-clones | full text |
| Listener sensitivity to clones (d′), Exp 1b / Exp 2 | 0.10 / 0.13 | lavan-2025-voice-clones-hyperrealism | full text |
| Realness rating, clones vs real (0–100) | 63.0 vs 63.8 (p = .682) | lavan-2025-voice-clones-hyperrealism | full text |
| Generic AI voices labelled human / d′ | 0.39 vs 0.72 for real / 2.93 | lavan-2025-voice-clones-hyperrealism | full text |
| Voice-clone accuracy; d′; at or below chance | 53.95% [52.64, 55.27]; .20; 35.47% of listeners | ivory-2026-faces-voices-domain | full text |
| Metacognitive AUROC, voices vs faces | .48 vs .59 | ivory-2026-faces-voices-domain | full text |
| TTS speech, single clip / pick-the-fake pair | 70.35% / 85.59% | mai-2023-speech-deepfakes | full text |
| Crowd-aggregated AUROC / out-of-domain detector | 95.51% / 25.31% | mai-2023-speech-deepfakes | full text |
| TTS vs voice-actor deepfakes identified | 72% vs 83% | groh-2024-political-speech | full text |
| Synthetic speech, Spanish and Japanese listeners | 59.11% | san-segundo-2026-audio-language-style | abstract |
| Pooled audio accuracy (k = 8) | 62.08% [38.23, 83.18] | diel-2024-deepfake-detection-meta | abstract |

## Key findings (citable) {#findings}

| ID | Finding | Confidence | Supporting sources |
|---|---|---|---|
| MR-0003.F1 | Commercial voice clones of unfamiliar speakers are accepted as the same identity as the real speaker in most pairs (median 83.3%; about 80% in the abstract). | Moderate (one verified large-N study) | barrington-2025-voice-clones |
| MR-0003.F2 | Listener sensitivity to modern voice clones is near chance: d′ 0.10–0.20, with 53.95–60.8% of clones correctly identified across three studies. | High | barrington-2025-voice-clones, lavan-2025-voice-clones-hyperrealism, ivory-2026-faces-voices-domain |
| MR-0003.F3 | Judgments of clones are biased toward "human/real" (c = −0.65; c = −.07), extending the truth default of MR-0001.F6 to voices. | Moderate | lavan-2025-voice-clones-hyperrealism, ivory-2026-faces-voices-domain |
| MR-0003.F4 | Clones reach realism parity with human voices but not hyperrealism (63.0 vs 63.8); generic, non-cloned AI voices remain easy to spot (d′ 2.93) and are rated more dominant. | Low | lavan-2025-voice-clones-hyperrealism |
| MR-0003.F5 | Detectability depends on synthesis method: older TTS is caught on ≈70% of single clips, state-of-the-art TTS less often than voice actors (72% vs 83%), and pooled audio accuracy (62.08%) has a CI spanning chance. | Moderate | mai-2023-speech-deepfakes, groh-2024-political-speech, diel-2024-deepfake-detection-meta, san-segundo-2026-audio-language-style |
| MR-0003.F6 | Comparison beats single judgment: choosing the fake from a pair reached 85.59% vs 70.35% for single clips, and same-sentence reference audio helped significantly; familiarisation (+3.84 points), extra time and repeated listening helped little. | Moderate | mai-2023-speech-deepfakes |
| MR-0003.F7 | Confidence carries no information about voice-clone accuracy (metacognitive AUROC .48), and voice-clone detection skill is not predicted by face-detection skill. | Low | ivory-2026-faces-voices-domain |
| MR-0003.F8 | Listeners rely on prosodic and extralinguistic cues (intonation, rhythm, fluency, pauses, breathing, laughter) and on delivery more than content; breathing cues predicted voice accuracy. | Moderate | san-segundo-2026-audio-language-style, ivory-2026-faces-voices-domain, groh-2024-political-speech |
| MR-0003.F9 | Aggregating many listeners is robust (AUROC 95.51%) where an out-of-domain automated detector failed (25.31%). | Low | mai-2023-speech-deepfakes |

Confidence (editorial, GRADE-inspired): **High** = consistent across ≥3 independent peer-reviewed studies, or one verified large-N study with internal replications; **Moderate** = fewer studies or heterogeneity; **Low** = one study or unverified details.

## Evidence table: accuracy by study {#evidence-table}

Accuracy here means the percentage of correct real-vs-synthetic judgments; chance is 50%.

| Source ID | Year | Modality | Stimuli | N | Headline result | Tier | Verification |
|---|---|---|---|---|---|---|---|
| barrington-2025-voice-clones | 2025 | audio | ElevenLabs clones of 220 speakers | 634 | same speaker 83.3%; clones caught 60.8% | B | full text |
| lavan-2025-voice-clones-hyperrealism | 2025 | audio | ElevenLabs clones + generic AI voices | 150 | clones d′ 0.10–0.13; generic d′ 2.93 | B | full text |
| ivory-2026-faces-voices-domain | 2026 | audio, face image | ElevenLabs clones; Firefly faces | 172 | voices 53.95% vs faces 65.48% | B | full text |
| mai-2023-speech-deepfakes | 2023 | audio | VITS TTS (EN, ZH) | 529 | 70.35% single; 85.59% paired | B | full text |
| groh-2024-political-speech | 2024 | audio, video, text | TTS vs voice actors | 2,215 | TTS 72% vs actors 83% | B | full text |
| san-segundo-2026-audio-language-style | 2026 | audio | Spanish/Japanese synthetic speech | 54 | 59.11% | B | abstract |
| diel-2024-deepfake-detection-meta | 2024 | audio subset | k = 8 | — | 62.08% [38.23, 83.18] | A | abstract |
| cooke-2024-coin-toss | 2025 | mixed incl. audio | 2023 generators | 1,276 | near 50%; foreign-language and single-modality items harder | B | abstract |

## Chart: correct real-vs-synthetic judgments by study {#chart-accuracy}

Text equivalent: VITS TTS pairs 85.59% and single clips 70.35% [mai-2023-speech-deepfakes](/data/citations.json#mai-2023-speech-deepfakes); TTS political speech 72% [groh-2024-political-speech](/data/citations.json#groh-2024-political-speech); pooled audio 62.08% [diel-2024-deepfake-detection-meta](/data/citations.json#diel-2024-deepfake-detection-meta); ElevenLabs clones identified 60.8% [barrington-2025-voice-clones](/data/citations.json#barrington-2025-voice-clones); Spanish/Japanese synthetic speech 59.11% [san-segundo-2026-audio-language-style](/data/citations.json#san-segundo-2026-audio-language-style); ElevenLabs clones 53.95% [ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain). The line marks 50% chance. Newer cloning tools sit closest to chance.

```mermaid
xychart-beta
    title "Human detection of synthetic speech by study (% correct)"
    x-axis ["TTS pair", "TTS single", "TTS speech", "Pooled audio", "Clones (B)", "ES/JA", "Clones (I)"]
    y-axis "Accuracy (%)" 0 --> 100
    bar [85.59, 70.35, 72, 62.08, 60.8, 59.11, 53.95]
    line [50, 50, 50, 50, 50, 50, 50]
```

## Theme 1 — Clones pass as the real speaker {#theme-identity}

- **Identity:** in 304 listeners' same/different judgments, real–clone pairs were judged the same speaker in a median 83.3% of trials, close to real–real pairs (100.0%) and far above different real speakers (11.1%); 26.6% of listeners said "same" on every real–clone trial [barrington-2025-voice-clones](/data/citations.json#barrington-2025-voice-clones).
- **Realness:** clones were labelled human about as often as real voices (0.58 vs 0.62; 0.70 vs 0.72) and rated as real (63.0 vs 63.8), so they are realistic but not hyperrealistic [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism). Compare AI faces in MR-0001.F2.
- **Scope:** every clone study used unfamiliar speakers and one commercial service (ElevenLabs).

## Theme 2 — Sensitivity is near chance and biased toward "human" {#theme-sensitivity}

- **d′ ≈ 0.1–0.2:** 0.10 and 0.13 across experiments [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism); .20, with 35.47% of listeners at or below chance [ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain); 21.0% at or below chance on clones [barrington-2025-voice-clones](/data/citations.json#barrington-2025-voice-clones).
- **Criterion:** c = −0.65 toward "human" for clones [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism); c = −.07 for voices [ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain).
- **Trust and dominance:** clones were not rated more trustworthy than real voices (58.9 vs 57.1, n.s.) but were rated more dominant (49.4 vs 44.0) [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism).

## Theme 3 — Synthesis method and language {#theme-method}

- **Older TTS:** VITS fakes 70.35% single clip (73% hit rate on fakes, 67.78% on genuine); English and Mandarin listeners similar (+1.79 points, n.s.) [mai-2023-speech-deepfakes](/data/citations.json#mai-2023-speech-deepfakes).
- **TTS vs human impersonators:** state-of-the-art TTS fakes were identified 72% of the time vs 83% for voice-actor fakes; audio alone 81% in Exp 1a [groh-2024-political-speech](/data/citations.json#groh-2024-political-speech).
- **Generic vs cloned AI voices:** generic designed voices were easy (d′ 2.93); clones were not [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism).
- **Language:** Spanish and Japanese listeners averaged 59.11%, better on authentic samples, with shared and language-specific cues [san-segundo-2026-audio-language-style](/data/citations.json#san-segundo-2026-audio-language-style). Foreign-language items lowered accuracy in a US sample [cooke-2024-coin-toss](/data/citations.json#cooke-2024-coin-toss).
- **Pooled:** audio 62.08% with a CI spanning chance (k = 8) [diel-2024-deepfake-detection-meta](/data/citations.json#diel-2024-deepfake-detection-meta).

## Theme 4 — What helps: comparison, crowds, unscripted speech {#theme-helps}

| Intervention or condition | Effect | Source (verification) |
|---|---|---|
| Pick the fake from a real/fake pair | 85.59% vs 70.35% single clip | mai-2023-speech-deepfakes (full text) |
| Same-sentence reference audio | significant gain (p < 0.001), rarely available in practice | mai-2023-speech-deepfakes (full text) |
| Familiarisation with 5 example clips | +3.84 points | mai-2023-speech-deepfakes (full text) |
| Extra time, repeated listening | little | mai-2023-speech-deepfakes (full text) |
| Aggregating many listeners | AUROC 95.51% | mai-2023-speech-deepfakes (full text) |
| Unscripted rather than scripted clips (exploratory, n = 30) | median 76.7% vs 56.7% | barrington-2025-voice-clones (full text) |

**Pattern:** procedures that give the listener a genuine comparison help far more than awareness or practice. No study has tested training against modern clones.

## Theme 5 — Confidence and cues {#theme-cues}

- **Confidence:** metacognitive AUROC .48 for voices (below chance) vs .59 for faces in the same people; face-detection d′ did not significantly predict voice d′ (β = .21, p = .053) [ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain).
- **Cues:** listeners relied on suprasegmental and extralinguistic cues (intonation, rhythm, fluency, pauses, speed, breathing, laughter) [san-segundo-2026-audio-language-style](/data/citations.json#san-segundo-2026-audio-language-style); breathing predicted voice accuracy [ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain); people relied more on how something was said than on what was said [groh-2024-political-speech](/data/citations.json#groh-2024-political-speech).
- **Machines:** in-domain speech detectors reached AUROC 100% but 25.31% out-of-domain, flagging genuine speech as fake [mai-2023-speech-deepfakes](/data/citations.json#mai-2023-speech-deepfakes).

## Moderators summary {#moderators}

| Moderator | Direction | Sources |
|---|---|---|
| Synthesis method | clones ≈ chance; older TTS ≈70%; generic AI voices easy | lavan-2025-voice-clones-hyperrealism, mai-2023-speech-deepfakes, groh-2024-political-speech |
| Paired vs single judgment | paired much easier | mai-2023-speech-deepfakes |
| Scripted vs unscripted speech | unscripted easier (exploratory) | barrington-2025-voice-clones |
| Listener language | EN vs ZH n.s.; ES vs JA share most cues; foreign language harder | mai-2023-speech-deepfakes, san-segundo-2026-audio-language-style, cooke-2024-coin-toss |
| Speaker familiarity | untested for clones; political-speech studies used well-known speakers (audio 81%) | groh-2024-political-speech |

## Contradictions and caveats {#caveats}

- Accuracies from 53.95% to 85.59% reflect task (single vs paired, generator, language), not disagreement.
- Mai et al. found confidence-weighted AUROC of 73.83%, so confidence carried some signal for older TTS; for clones it did not [ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain).
- All studies warned listeners that fakes were present; real-world detection is likely lower.

## Evidence gaps {#gaps}

1. No peer-reviewed study of clones of **familiar** voices (a colleague, a family member, an executive) was found.
2. No study tests telephone-band audio, background noise or live two-way conversation (vishing conditions).
3. No study of real-time voice conversion in calls.
4. No training study against modern clones, and no retention data.
5. One cloning service (ElevenLabs) dominates the evidence.

## Implications (editorial inference, not study findings) {#implications}

- **Do not authenticate by voice.** A call that "sounds like" the person is not evidence (F1, F2, F7).
- **Use callbacks and out-of-band checks** rather than listening harder; comparison with a known-genuine channel is the only listener-side aid with a large effect (F6).
- **Expect a "human" default** in staff judgments (F3).
- **Do not rely on one off-the-shelf audio detector** outside its training domain (F9).

## Answers to common questions {#faq}

**Q: Can people tell a cloned voice from the real person?**
A: Mostly not, for unfamiliar speakers: same speaker in a median 83.3% of pairs [barrington-2025-voice-clones](/data/citations.json#barrington-2025-voice-clones); d′ ≈ 0.1 [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism) (MR-0003.F1, F2).

**Q: Are AI voices more convincing than real ones?**
A: Not yet: clones are rated as real as human voices, not more (MR-0003.F4).

**Q: What helps people spot a fake voice?**
A: A genuine reference to compare against (85.59% vs 70.35%) [mai-2023-speech-deepfakes](/data/citations.json#mai-2023-speech-deepfakes); practice and extra time help little (MR-0003.F6).

**Q: Does it help to know the speaker?**
A: Untested for modern clones; this is the main gap (see [Gaps](#gaps)).

## Method of this metareport {#method}

- **Type:** narrative synthesis; pooled audio value quoted from Diel et al. (2024).
- **Inclusion:** peer-reviewed studies of human listeners judging synthetic or cloned speech against real speech.
- **Origin:** split from MR-0001 on 2026-10-19 because MR-0001 exceeded its page-size limit; MR-0001.F12 and F15 remain valid and cross-reference this report.
- **Search limits:** discovery via OpenAlex, PubMed and reference lists, without general web search. 2025–2026 coverage is incomplete.

## Verification status {#verification}

| Status | Count | Sources |
|---|---|---|
| `verified-fulltext` | 5 | barrington-2025-voice-clones, lavan-2025-voice-clones-hyperrealism, ivory-2026-faces-voices-domain, mai-2023-speech-deepfakes, groh-2024-political-speech |
| `verified-abstract` | 3 | san-segundo-2026-audio-language-style, diel-2024-deepfake-detection-meta, cooke-2024-coin-toss |
| `verified-doi` | 0 | — |
| `unverified` | 0 | — |

Values from `unverified` records are marked "≈" or "unverified" in this report. Check them against the primary source before high-stakes use. Each source page states exactly what was checked and how.

## References {#references}

Source pages contain the full structured records: design, N, results, limitations and verification.

- [barrington-2025-voice-clones](/data/citations.json#barrington-2025-voice-clones) · [record](/sources/barrington-2025-voice-clones.md) · Barrington, S., et al. (2025). People are poorly equipped to detect AI-powered voice clones. *Scientific Reports*, 15, 11004. · https://doi.org/10.1038/s41598-025-94170-3
- [cooke-2024-coin-toss](/data/citations.json#cooke-2024-coin-toss) · [record](/sources/cooke-2024-coin-toss.md) · Cooke, D., et al. (2025). As good as a coin toss: Human detection of AI-generated images, videos, audio, and audiovisual stimuli. *Communications of the ACM*, 68, 100-109. · https://doi.org/10.1145/3729417
- [diel-2024-deepfake-detection-meta](/data/citations.json#diel-2024-deepfake-detection-meta) · [record](/sources/diel-2024-deepfake-detection-meta.md) · Diel, A., et al. (2024). Human performance in detecting deepfakes: A systematic review and meta-analysis of 56 papers. *Computers in Human Behavior Reports*, 16, 100538. · https://doi.org/10.1016/j.chbr.2024.100538
- [groh-2024-political-speech](/data/citations.json#groh-2024-political-speech) · [record](/sources/groh-2024-political-speech.md) · Groh, M., et al. (2024). Human detection of political speech deepfakes across transcripts, audio, and video. *Nature Communications*, 15, 7629. · https://doi.org/10.1038/s41467-024-51998-z
- [ivory-2026-faces-voices-domain](/data/citations.json#ivory-2026-faces-voices-domain) · [record](/sources/ivory-2026-faces-voices-domain.md) · Ivory, M. R., et al. (2026). Human detection of AI-generated faces and voices is not domain-general. *Scientific Reports*, 16, 28639. · https://doi.org/10.1038/s41598-026-59364-3
- [lavan-2025-voice-clones-hyperrealism](/data/citations.json#lavan-2025-voice-clones-hyperrealism) · [record](/sources/lavan-2025-voice-clones-hyperrealism.md) · Lavan, N., et al. (2025). Voice clones sound realistic but not (yet) hyperrealistic. *PLOS ONE*, 20(9), e0332692. · https://doi.org/10.1371/journal.pone.0332692
- [mai-2023-speech-deepfakes](/data/citations.json#mai-2023-speech-deepfakes) · [record](/sources/mai-2023-speech-deepfakes.md) · Mai, K. T., et al. (2023). Warning: Humans cannot reliably detect speech deepfakes. *PLOS ONE*, 18(8), e0285333. · https://doi.org/10.1371/journal.pone.0285333
- [san-segundo-2026-audio-language-style](/data/citations.json#san-segundo-2026-audio-language-style) · [record](/sources/san-segundo-2026-audio-language-style.md) · San Segundo, E., et al. (2026). Human perception of audio deepfakes: the role of language and speaking style. *Speech Communication*, 185, 103487. · https://doi.org/10.1016/j.specom.2026.103487