Voice clones sound realistic but not (yet) hyperrealistic
Source record
lavan-2025-voice-clones-hyperrealism· 2025 · PLOS ONE
Evidence tier: B (peer-reviewed-journal) · Verification: verified-fulltext
Bottom line: Listeners could not tell voice clones from the real speakers (d' 0.10-0.13) and rated them as real as human voices, but clones were not judged more human than humans (no hyperrealism); AI voices were rated more dominant.
Citation
Nadine Lavan, Mairi Irvine, Victor Rosi, Carolyn McGettigan (2025). Voice clones sound realistic but not (yet) hyperrealistic. PLOS ONE, 20(9), e0332692. https://doi.org/10.1371/journal.pone.0332692
Verification entry (identifiers, resolvers, claims with source locations): /data/citations.json#lavan-2025-voice-clones-hyperrealism
Study design
| Field | Value |
|---|---|
| Design | Three online experiments: generic AI voices vs human (Exp1a), voice clones vs the same humans (Exp1b), both together (Exp2); realness, human-vs-AI classification, trustworthiness and dominance ratings |
| Modality | audio |
| Stimuli | 40 VCTK speakers; 40 ElevenLabs Instant Voice Cloning clones (≈4 min of training audio each); 40 ElevenLabs Voice Design generic voices; one Rainbow Passage sentence each, artefact-screened |
| Participants (N) | 150 |
| N detail | n=50 per experiment; UK native English speakers via Prolific; not preregistered |
| Topics | human-detection, voice, voice-clone, hyperrealism, trust |
Results
| Metric | Value | Note |
|---|---|---|
| Exp1b proportion labelled human: clones vs real | 0.58 vs 0.62 | OR = 0.84, p = .239; d' = 0.10, n.s. |
| Exp2 proportion labelled human: clones vs real | 0.70 vs 0.72 | d' clones 0.13; generic AI voices 0.39 vs 0.72, d' 2.93 |
| Exp2 realness ratings: clones vs real | 63.0 vs 63.8 | p = .682 |
| Dominance: clones vs real | 49.4 vs 44.0 | p = .015; generic AI 58.1 vs 40.4 |
| Trustworthiness: clones vs real | 58.9 vs 57.1 | n.s.; generic AI voices 57.9 vs 53.8, p = .002 |
Moderators
- Generic (non-cloned) AI voices were easy to spot; clones were not
- Criterion biased toward 'human' for clones (c = -0.65 in Exp2)
Limitations
- Unfamiliar voices only
- Artefact-screened, idealised stimuli
- One sentence per voice
- Not preregistered
- Context effects across experiments
Cited in
- MR-0001: Human Detection of Deepfakes: Accuracy, Bias, Confidence, and Training
- MR-0002: The Uncanny Valley and Synthetic Media: Does It Exist, What Causes It, and Have Deepfakes Crossed It?
- MR-0003: Voice Clones and Synthetic Speech: How Well Do People Detect Them?
Verification
| Field | Value |
|---|---|
| Status | verified-fulltext |
| Compiler confidence | high |
| Checked on | 2026-10-12 |
| Via | journals.plos.org full text |