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).

TL;DR

• 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.
• Near-chance sensitivity: d′ 0.10–0.13 for ElevenLabs clones lavan-2025-voice-clones-hyperrealism; 53.95% correct (d′ .20) ivory-2026-faces-voices-domain; clones identified 60.8% barrington-2025-voice-clones.
• Bias toward "human": criterion c = −0.65 for clones lavan-2025-voice-clones-hyperrealism; c = −.07 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.
• Older and generic voices are easier: VITS TTS 70.35% single clip mai-2023-speech-deepfakes; generic AI voices d′ 2.93 lavan-2025-voice-clones-hyperrealism; TTS fakes caught 72% vs 83% for voice actors 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.
• Confidence is no guide: metacognitive AUROC .48 for voices ivory-2026-faces-voices-domain.

Key numbers

Key findings (citable)

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

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

Chart: correct real-vs-synthetic judgments by study

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

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

• 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.
• 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. 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"

• d′ ≈ 0.1–0.2: 0.10 and 0.13 across experiments lavan-2025-voice-clones-hyperrealism; .20, with 35.47% of listeners at or below chance ivory-2026-faces-voices-domain; 21.0% at or below chance on clones barrington-2025-voice-clones.
• Criterion: c = −0.65 toward "human" for clones lavan-2025-voice-clones-hyperrealism; c = −.07 for voices 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.

Theme 3 — Synthesis method and language

• 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.
• 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.
• Generic vs cloned AI voices: generic designed voices were easy (d′ 2.93); clones were not 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. Foreign-language items lowered accuracy in a US sample cooke-2024-coin-toss.
• Pooled: audio 62.08% with a CI spanning chance (k = 8) diel-2024-deepfake-detection-meta.

Theme 4 — What helps: comparison, crowds, unscripted speech

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

• 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.
• Cues: listeners relied on suprasegmental and extralinguistic cues (intonation, rhythm, fluency, pauses, speed, breathing, laughter) san-segundo-2026-audio-language-style; breathing predicted voice accuracy ivory-2026-faces-voices-domain; people relied more on how something was said than on what was said 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.

Moderators summary

Contradictions and 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.
• All studies warned listeners that fakes were present; real-world detection is likely lower.

Evidence 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)

• 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

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; d′ ≈ 0.1 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; 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).

Method of this metareport

• 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

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

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

• barrington-2025-voice-clones · record · 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 · record · 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 · record · 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 · record · 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 · record · 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 · record · 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 · record · 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 · record · 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