When 172 people were asked to sort real from AI-generated voices, they got it right just 53.95% of the time, barely above a coin flip, while the same participants classified faces correctly 65.48% of the time (Ivory et al., Scientific Reports 2026, in press). Clones of real people fool listeners even on identity: matching accuracy fell from 86.98% for natural voices to 68.94% for deepfakes (Roswandowitz et al., Communications Biology 2024). Across 1,212 participants and three widely used deepfake datasets, human accuracy averaged 73% (Warren et al., ACM CCS 2024), yet the classic uncanny valley, the dip in comfort when something is almost but not quite human, barely shows up for synthetic speech on its own. We aggregated data from PLOS One, Scientific Reports, Communications Biology, eNeuro, the ISCA Archive, Frontiers, ACM CCS, i-Perception and two 2026 industry surveys. For the listening-test benchmarks behind modern text-to-speech, see our synthetic speech naturalness and MOS statistics, and for the pitch and stress errors that still give synthetic voices away, our neural TTS prosody error statistics.
TL;DR
- Listeners classified real vs synthetic voices correctly 53.95% of the time, vs 65.48% for faces (Ivory et al., Scientific Reports 2026).
- Confidence did not track accuracy for voices: AUROC .48, where .5 means no insight (Ivory et al., Scientific Reports 2026).
- Voice clones were labelled human at 0.70, real recordings at 0.72 and generic AI voices at 0.39 (Lavan et al., PLOS One 2025).
- Mean human accuracy across three deepfake datasets was 73% for 1,212 participants (Warren et al., ACM CCS 2024).
- Identity matching fell from 86.98% for natural voices to 68.94% for deepfakes (Roswandowitz et al., Communications Biology 2024).
- The more human-like a TTS voice, the more pleasant it was rated: r(203) = .48, p < .001 (Ross, Corley and Lai, Speech Prosody 2024).
- Mismatched face and voice pairings were rated eerier than matched ones, t(47) = 6.042, p < 0.001 (Mitchell et al., i-Perception 2011).
- A human voice lifted a virtual character’s likability to 3.75 vs 2.99 for a robot voice (Choi et al., IEEE ISMAR-Adjunct 2023).
- After learning the voiceover they heard was AI, 25% of radio listeners felt more favorable and 20% less favorable (Crowd React Media 2026).
- Example-based familiarization raised deepfake detection by just 3.84% on average (Mai et al., PLOS ONE 2023).
- 67% of adults in the US, UK and DACH region worry about having their voice cloned (Malwarebytes 2026, vendor survey).
1. Detection: How Often Listeners Tell AI Voices From Human Ones
The headline is not that people fail, it is that the failure is specific to hearing. The same participants who beat chance comfortably on faces scored 53.95% on voices (95% CI = [52.64, 55.27]) in a preregistered Lancaster University study that tested both senses side by side (Ivory et al., Scientific Reports, in press). Being good at spotting fake faces said nothing about being good at spotting fake voices: the authors found no general, cross-sense detection ability.
Task format moves the numbers more than listener skill does. When UCL researchers played one clip at a time, 529 listeners were right 70.35% of the time, but when a genuine and a deepfake clip were heard together, the fake was recognized in 85.59% of trials (Mai et al., PLOS ONE 2023). Real life rarely offers the comparison clip. Dataset age matters too: in the University of Florida’s study of 1,212 participants, accuracy ranged from 65% on the Wavefake set to 81% on FakeAVCeleb.
| Metric | Value | Source |
|---|---|---|
| Real vs synthetic voices classified correctly (172 participants, 40 real and 40 synthetic voices) | 53.95% (95% CI = [52.64, 55.27]) | Ivory et al., Scientific Reports 2026 (in press) |
| Real vs AI-generated faces classified correctly by the same participants | 65.48% (95% CI = [63.75, 67.21]) | Ivory et al., Scientific Reports 2026 (in press) |
| Correct classifications when judging one clip at a time (n = 529, English and Mandarin) | 70.35% | Mai et al., PLOS ONE 2023 |
| Deepfake recognized when heard alongside a genuine clip (two-clip task) | 85.59% of trials | Mai et al., PLOS ONE 2023 |
| Mean human accuracy across three deepfake datasets (1,212 participants) | 73% | Warren et al., ACM CCS 2024 |
| Mean accuracy on the hardest dataset (Wavefake) vs the easiest (FakeAVCeleb) | 65% vs 81% | Warren et al., ACM CCS 2024 |
| AI voice clones correctly identified as AI (Prolific sample of 604 participants across two studies) | about 60% | Barrington, Cooper and Farid, Scientific Reports 2025 |
| Listener sensitivity (d prime) separating generic AI voices from human voices vs voice clones from human voices (n = 50) | 2.93 vs 0.13 | Lavan et al., PLOS One 2025 |
Context note: d prime (d’) is a sensitivity score from signal detection theory in which 0 means the listener cannot separate the two categories at all. The Mai et al. deepfakes came from relatively old algorithms, as the authors point out; most recent available data for the one-clip vs two-clip comparison: Mai et al., PLOS ONE 2023.
2. Voice Clones: When the Copy Passes for the Person
A generic AI voice still sounds like a product. A clone of a specific person does not, and that is where human hearing gives up. Researchers at Queen Mary University of London built their clones from around 4 minutes of recordings per speaker with a commercial cloning tool, and listeners labelled those clones human at a rate of 0.70, against 0.72 for the real recordings (Lavan et al., PLOS One 2025). The authors’ verdict is precise: clones sound real, but not yet “hyperreal” the way AI faces can look more real than real ones. How little audio a clone now needs is tracked in our zero-shot voice cloning reference audio statistics.
Identity is the second test, and clones pass it most of the time. In a UC Berkeley study, participants judged an AI voice to have the same identity as its real counterpart approximately 80% of the time (Barrington, Cooper and Farid, arXiv v2, published in Scientific Reports 2025). In a University of Zurich brain-imaging study that used German two-word sentences (the study stimulus included ‘Er fällt’, German for ‘He falls’), identity matching dropped from 86.98% for natural voices to 68.94% for deepfakes (Roswandowitz et al., Communications Biology 2024).
| Metric | Value | Source |
|---|---|---|
| Share of voices labelled ‘human’: voice clones vs real human recordings vs generic AI voices (Experiment 2, n = 50 UK adults) | 0.70 vs 0.72 vs 0.39 | Lavan et al., PLOS One 2025 |
| Mean realness rating: generic AI voices vs voice clones vs real human recordings | 47.4 vs 63.0 vs 63.8 | Lavan et al., PLOS One 2025 |
| AI clone judged to have the same identity as the real speaker | approximately 80% of the time | Barrington, Cooper and Farid, Scientific Reports 2025 |
| Correlation between clip duration and accuracy of naturalness judgments | Spearman ρ = 0.245 | Barrington, Cooper and Farid, Scientific Reports 2025 |
| Identity-matching accuracy: pairs of natural voices vs a deepfake paired with a natural voice (n = 25) | 86.98% vs 68.94% | Roswandowitz et al., Communications Biology 2024 |
| Deepfake trials: misses vs false alarms | 28.64% vs 29.86% | Roswandowitz et al., Communications Biology 2024 |
Context note: the Queen Mary team ran three experiments with 50 UK participants each, aged 18 to 65 and recruited on Prolific. In the Zurich study, a miss means a deepfake was accepted as the real speaker, and a false alarm means a correct identity was rejected; the near-equal rates show listeners were guessing in both directions rather than leaning one way.
3. Is There an Uncanny Valley for Voices?
The visual uncanny valley predicts that comfort falls just before an artificial face becomes fully human. For synthetic speech, most data point the other way: the closer to human, the better. In a Frontiers in Neurorobotics study, 95 adults rated two synthetic voices (IBM Watson and the Sophia robot) and a human voice, and higher human-likeness predicted lower eeriness for both synthetic voices. An Edinburgh team found that human-likeness and pleasantness were positively correlated, r(203) = .48, p < .001, but no TTS voice reached 50% human-likeness, so the top of the curve, where a valley would sit, was never tested (Ross, Corley and Lai, Speech Prosody 2024).
When a valley does appear, it comes from deviation rather than from artificiality. Diel and Lewis found a vocal uncanny valley driven by distorted and pathological voices, and concluded that “synthetic voices successfully escape it” (Computers in Human Behavior Reports 2024). For anyone choosing a voice with an AI voice generator, that is a practical rule: an unusual, strained or glitchy delivery is riskier than a plainly synthetic one.
| Metric | Value | Source |
|---|---|---|
| Human-likeness rating: most human-like TTS voice (exaggerated pitch range) vs human recording (205 listeners) | 46.1 vs 71.5 | Ross, Corley and Lai, Speech Prosody 2024 |
| Eeriness rating: same TTS voice vs human recording | 40.4 vs 28.9 | Ross, Corley and Lai, Speech Prosody 2024 |
| Correlation between human-likeness and pleasantness | r(203) = .48, p < .001 | Ross, Corley and Lai, Speech Prosody 2024 |
| Human-likeness predicting eeriness for two synthetic voices (95 adults) | β = −0.64 and β = −0.66, p < 0.001 | Kühne, Fischer and Zhou, Frontiers in Neurorobotics 2020 |
| Voices in a likability vs human-likeness study (25 listeners): German male voices from 5 synthesis approaches plus a human control | 13 | Baird et al., Interspeech 2018 |
| Participants in two experiments locating a vocal uncanny valley in deviating voices | n = 50 and n = 35 | Diel and Lewis, Computers in Human Behavior Reports 2024 |
| Participants in a voice assistant study where human-like appearance raised eeriness | 34 | Zhao, Chen and Zhao, Human Factors and Ergonomics in Manufacturing 2026 |
| Exposures before liking of an unfamiliar human voice rose (spoken sentences, 238 university students) | 8 | Mavromati, Erdinch and Paterson, PsyArXiv preprint 2025 |
Context note: Baird et al. found that likability rose with human-likeness, the opposite of the visual uncanny valley (Interspeech 2018); most recent available data for that comparison across synthesis generations, and the Kühne et al. study dates from 2020. In the voice assistant study, the eeriness came from a human-like appearance, not from the voice. The preprint by Mavromati and colleagues tested real human voices: those heard 2 or 5 times were liked less than never-heard voices, an inverted ‘U’ the authors call an uncanny valley of ‘familiarly unfamiliar’ voices.
4. Face-Voice Mismatch: Where the Valley Actually Shows Up
If voices alone rarely fall into the valley, combinations do. The foundational test paired a robot or a human figure with a matching or mismatching voice, and the two mismatched pairings were rated eerier than the two matched ones, t(47) = 6.042, p < 0.001 (Mitchell et al., i-Perception 2011). The human figure with a synthetic voice was the eeriest of the four conditions. The design lesson the authors drew still holds: match the human realism of the face to that of the voice.
Newer work shows the voice usually dominates the impression. In a virtual reality study, a human voice raised a character’s likability and perceived safety, while appearance itself had no significant effect on likability (Choi et al., IEEE ISMAR-Adjunct 2023). With the Furhat robot head, the face-voice setting changed first-impression warmth, but the effect was no longer significant after participants interacted with the robot (Huang and Moore, Frontiers in Robotics and AI 2024).
| Metric | Value | Source |
|---|---|---|
| Eeriness rating: human figure with synthetic voice vs human figure with human voice (48 US-born participants) | M = 0.19 vs M = −1.10 | Mitchell et al., i-Perception 2011 |
| Eeriness rating: robot figure with human voice vs robot figure with synthetic voice | M = −0.10 vs M = −0.60 | Mitchell et al., i-Perception 2011 |
| Mismatched vs matched face and voice pairings (paired t-test) | t(47) = 6.042, p < 0.001 | Mitchell et al., i-Perception 2011 |
| Likability of a virtual character: human voice vs robot voice (21 participants) | 3.75 vs 2.99 | Choi et al., IEEE ISMAR-Adjunct 2023 |
| Perceived human-likeness (anthropomorphism) of a virtual character: human voice vs robot voice | 3.18 vs 2.58 | Choi et al., IEEE ISMAR-Adjunct 2023 |
| First-impression warmth: child-like vs adult-like vs robot-like face and voice settings (52 participants, 156 samples) | 3.37 vs 2.60 vs 2.06 | Huang and Moore, Frontiers in Robotics and AI 2024 |
| Effect of the face and voice setting on warmth before interaction | F(2,147) = 16.75, p < 0.001 | Huang and Moore, Frontiers in Robotics and AI 2024 |
Context note: in the Furhat study some participants described the child-like setting as mismatched (‘a more child-like sound but less child-like look’). A 2x2 virtual-agent experiment by Alimardani et al. (IVA 2024) found that a voice-appearance mismatch changed perceived anthropomorphism but did not change how often people followed the agent’s advice. The Mitchell et al. study is from 2011; most recent available data for this exact face-voice design.
5. Social Traits and the Disclosure Effect
Listeners do not just judge whether a voice is real; they read character into it, and synthetic voices are read differently. In the Queen Mary trait experiments, generic AI voices were rated more dominant and slightly more trustworthy than human voices. The Zurich team found the reverse for its deepfakes, which participants rated “significantly less natural, likable, and trustworthy” than the natural voices (Roswandowitz et al., Communications Biology 2024). Stimulus quality decides which way the traits fall.
In a blind test closer to real media, quality stopped being the issue and disclosure took over. After the reveal, 48% of listeners who had heard the human voiceover felt more favorable, against 25% of those who had heard the AI voice (Crowd React Media AI Voiceover Study 2026). Before anyone was told anything, the two voices scored almost the same. Flat delivery is the other weak spot: approximately 68% of users in a no-emotion condition criticized a conversational agent’s voice tone (Cihodaru-Ștefanache and Podina, Frontiers in Psychology 2025).
| Metric | Value | Source |
|---|---|---|
| Dominance rating: generic AI voices vs human voices (n = 50) | 58.1 vs 40.4 (p < .001) | Lavan et al., PLOS One 2025 |
| Trustworthiness rating: generic AI voices vs human voices | 57.9 vs 53.8 (p = .002) | Lavan et al., PLOS One 2025 |
| Overall appeal of the first clip: AI voiceover vs human voiceover (1,326 US radio listeners aged 18 to 45) | 61% vs 60% | Crowd React Media AI Voiceover Study 2026 |
| Voiceover rated ‘funny’: human voice vs AI voice (the only significant gap) | 33% vs 26% | Crowd React Media AI Voiceover Study 2026 |
| After the reveal, listeners of the AI voice: more favorable vs less favorable | 25% vs 20% | Crowd React Media AI Voiceover Study 2026 |
| After the reveal, listeners of the human voice: more favorable vs less favorable | 48% vs 4% | Crowd React Media AI Voiceover Study 2026 |
| Users in a no-emotion condition who criticized a conversational agent’s voice tone | approximately 68% | Cihodaru-Ștefanache and Podina, Frontiers in Psychology 2025 |
Context note: the Frontiers in Psychology review screened its way to 29 studies of embodied conversational agents (animated on-screen characters that talk with users) and rated most of them weak or moderate in methodological quality. The radio test was run by a market research division of Harker Bos Group and checked significance with a two-proportion z-test at 95% confidence; listeners were not told which voice they heard until after rating it.
6. Training, Confidence and the Listening Brain
Short training barely helps, and listeners cannot tell when they are wrong. For voices, metacognitive insight measured as AUROC was .48, just under the .5 that means confidence carries no information about accuracy, compared with .59 for faces in the same Lancaster study (see section 1). Familiarization with example deepfakes barely moves accuracy (see the table). When listeners were unsure, the University of Florida team found, they leaned toward believing the audio was human (Warren et al., ACM CCS 2024), which is exactly the bias voice-cloning fraud relies on; see our voice cloning family fraud statistics.
The brain, however, registers more than the listener reports. After a 12-minute training session, 30 native Mandarin speakers showed neural differences between human and AI speech at about 55, 210 and 455 milliseconds, yet their behavioral discrimination did not improve significantly (Yang et al., eNeuro 2026). In the Zurich fMRI study, the auditory cortex tracked acoustic differences between deepfake and natural voices, while the nucleus accumbens, part of the reward system, was less active when a deepfake was compared with a natural voice (Roswandowitz et al., Communications Biology 2024).
| Metric | Value | Source |
|---|---|---|
| Metacognitive insight (AUROC) for voices vs faces, where .5 means no insight | .48 vs .59 | Ivory et al., Scientific Reports 2026 (in press) |
| Detection gain from familiarization with deepfake examples | 3.84% on average | Mai et al., PLOS ONE 2023 |
| Sensitivity (d prime) for human speech before vs after a 12-minute training session (30 participants) | 0.383 vs 0.536 | Yang et al., eNeuro 2026 |
| Sensitivity (d prime) for fine-tuned AI clones before vs after the same training | 0.266 vs 0.341 | Yang et al., eNeuro 2026 |
| Main effect of the training session on sensitivity | F(1,29) = 3.918, p = 0.057 | Yang et al., eNeuro 2026 |
| Latencies of neural differentiation between human and AI speech after training | about 55 ms, 210 ms and 455 ms | Yang et al., eNeuro 2026 |
| Adults worried about having their voice cloned (1,500 adults in the US, UK, Austria, Germany and Switzerland) | 67% | Malwarebytes AI Scams Report 2026 (vendor survey) |
| Adults who received a voice-cloned call from someone they knew in the past 12 months | 16% | Malwarebytes AI Scams Report 2026 (vendor survey) |
Context note: AUROC (area under the receiver operating characteristic curve) here measures whether a listener’s confidence is higher on correct answers than on wrong ones. The eNeuro sentences came from a Mandarin speech perception test, and the AI voices were made with an open-source cloning model. The Malwarebytes survey was fielded online from March 1 to March 17, 2026 by an independent research consultant; it is a security vendor’s own survey, so treat it as context, not as a perception measurement.
Summary: Uncanny Valley in Synthetic Voices by the Numbers
| Metric | Value | Source |
|---|---|---|
| Real vs synthetic voices classified correctly | 53.95% | Ivory et al., Scientific Reports 2026 (in press) |
| Mean human accuracy across three deepfake datasets (1,212 participants) | 73% | Warren et al., ACM CCS 2024 |
| Mean accuracy on the hardest vs easiest deepfake dataset | 65% vs 81% | Warren et al., ACM CCS 2024 |
| Deepfake detection: one clip at a time vs alongside a genuine clip | 70.35% vs 85.59% | Mai et al., PLOS ONE 2023 |
| AI clone judged to have the same identity as the real speaker | approximately 80% | Barrington, Cooper and Farid, Scientific Reports 2025 |
| Share labelled ‘human’: voice clones vs real recordings vs generic AI voices | 0.70 vs 0.72 vs 0.39 | Lavan et al., PLOS One 2025 |
| Identity matching: natural voices vs deepfakes | 86.98% vs 68.94% | Roswandowitz et al., Communications Biology 2024 |
| Correlation between human-likeness and pleasantness for TTS voices | r(203) = .48, p < .001 | Ross, Corley and Lai, Speech Prosody 2024 |
| Human-likeness: most human-like TTS voice vs human recording | 46.1 vs 71.5 | Ross, Corley and Lai, Speech Prosody 2024 |
| Eeriness: most human-like TTS voice vs human recording | 40.4 vs 28.9 | Ross, Corley and Lai, Speech Prosody 2024 |
| Human-likeness predicting eeriness for two synthetic voices | β = −0.64 and β = −0.66 | Kühne, Fischer and Zhou, Frontiers in Neurorobotics 2020 |
| Mismatched vs matched face and voice pairings | t(47) = 6.042, p < 0.001 | Mitchell et al., i-Perception 2011 |
| Eeriness: human figure with synthetic voice vs human figure with human voice | M = 0.19 vs M = −1.10 | Mitchell et al., i-Perception 2011 |
| Likability of a virtual character: human voice vs robot voice | 3.75 vs 2.99 | Choi et al., IEEE ISMAR-Adjunct 2023 |
| First-impression warmth: child-like vs adult-like vs robot-like settings | 3.37 vs 2.60 vs 2.06 | Huang and Moore, Frontiers in Robotics and AI 2024 |
| Overall appeal: AI voiceover vs human voiceover | 61% vs 60% | Crowd React Media AI Voiceover Study 2026 |
| Users who criticized an emotionless conversational agent’s voice tone | approximately 68% | Cihodaru-Ștefanache and Podina, Frontiers in Psychology 2025 |
| Effect of a 12-minute training session on sensitivity to AI speech | F(1,29) = 3.918, p = 0.057 | Yang et al., eNeuro 2026 |
Methodology and Sources
- Ivory, McGettigan, Nuttall and Nightingale, “Human Detection of AI-Generated Faces and Voices is Not Domain-General”, Scientific Reports 2026, in press, Lancaster University repository
- Lavan et al., “Voice clones sound realistic but not (yet) hyperrealistic”, PLOS One 2025, PubMed Central
- Barrington, Cooper and Farid, “People are poorly equipped to detect AI-powered voice clones”, Scientific Reports 2025, PubMed Central and arXiv v2
- Mai, Bray, Davies and Griffin, “Warning: Humans cannot reliably detect speech deepfakes”, PLOS ONE 2023, PLOS ONE
- Warren et al., “Better Be Computer or I’m Dumb: A Large-Scale Evaluation of Humans as Audio Deepfake Detectors”, ACM CCS 2024, University of Florida PDF
- Roswandowitz et al., “Cortical-striatal brain network distinguishes deepfake from real speaker identity”, Communications Biology 2024, PubMed Central
- Yang, Jiang, Bai, Ni and Teng, “Short-Term Perceptual Training Modulates Neural Responses to Deepfake Speech but Does Not Improve Behavioral Discrimination”, eNeuro 2026, PubMed Central
- Ross, Corley and Lai, “Is there an uncanny valley for speech? Investigating listeners’ evaluations of realistic TTS voices”, Speech Prosody 2024, ISCA Archive
- Kühne, Fischer and Zhou, “The Human Takes It All: Humanlike Synthesized Voices Are Perceived as Less Eerie and More Likable”, Frontiers in Neurorobotics 2020, Frontiers
- Baird et al., “The Perception and Analysis of the Likeability and Human Likeness of Synthesized Speech”, Interspeech 2018, ISCA Archive
- Diel and Lewis, “Deviation from typical organic voices best explains a vocal uncanny valley”, Computers in Human Behavior Reports 2024, ScienceDirect
- Zhao, Chen and Zhao, “Does the Uncanny Valley Effect Exist When Voice Assistants Have Minds and Human-Like Appearances?”, Human Factors and Ergonomics in Manufacturing and Service Industries 2026, Wiley via DOI
- Mavromati, Erdinch and Paterson, “The uncanny valley of familiarly unfamiliar voices”, PsyArXiv preprint 2025, OSF
- Mitchell et al., “A Mismatch in the Human Realism of Face and Voice Produces an Uncanny Valley”, i-Perception 2011, PubMed Central
- Choi, Koilias, Volonte, Kao and Mousas, “Exploring the Appearance and Voice Mismatch of Virtual Characters”, IEEE ISMAR-Adjunct 2023, PDF
- Huang and Moore, “Freedom comes at a cost?: An exploratory study on affordances’ impact on users’ perception of a social robot”, Frontiers in Robotics and AI 2024, Frontiers
- Alimardani, de Roode, Vaitonyte and Louwerse, “Effect of a Virtual Agent’s Appearance and Voice on Uncanny Valley and Trust in Human-Agent Collaboration”, IVA 2024, ACM via DOI
- Cihodaru-Ștefanache and Podina, “The uncanny valley effect in embodied conversational agents: a critical systematic review”, Frontiers in Psychology 2025, Frontiers
- Crowd React Media (Harker Bos Group), AI Voiceover Study 2026, Crowd React Media, also covered by Radio Ink
- Malwarebytes, AI Scams Report 2026 (vendor survey of 1,500 adults), Malwarebytes
- Data watch: the Lancaster face-and-voice study is cited from the accepted version in the university repository, so figures could shift slightly in the final journal version. The UC Berkeley clone study prints its headline percentages as images in the journal version, so the approximately 80% and about 60% figures are taken from the arXiv v2 text, which matches the published abstract. The Queen Mary baseline moved between experiments (listeners labelled 81% of human voices ‘human’ in Experiment 1a, 62% in Experiment 1b and 72% in Experiment 2), so clones and human voices should be compared within one experiment, never across them. The two 73% figures from the UCL (2023) and University of Florida (2024) studies measure different tasks and datasets and are not a replication of each other. The radio blind test did not find a statistically significant difference in how often the two voices were identified, and its exact identification shares appear only in a chart. No published test here reached TTS voices rated close to fully human-like, so the top of the vocal uncanny valley curve remains untested. Benchmark naturalness scores (MOS and CMOS) are deliberately left to our MOS roundup.
Last updated: November 8, 2026. We update this roundup quarterly, and the next refresh is expected when the final version of the Lancaster face-and-voice detection study is published in Scientific Reports.