In a field test of commercial voice stress analysis with jail arrestees, two popular programs flagged only about 15% of the lies that arrestees told about recent drug use, and their overall accuracy was about 50% (National Institute of Justice, Voice Stress Analysis field test 2008). Human listeners do little better: across 24,483 judges, people average 54% correct lie-truth judgments, barely above a coin toss (Personality and Social Psychology Review, Accuracy of Deception Judgments 2006). The National Research Council found that the research offers “little or no scientific basis for the use of the computer voice stress analyzer” (National Research Council, The Polygraph and Lie Detection 2003). Even so, the main manufacturer still claims thousands of police customers, and listening-only studies from 2025 and 2026 are reopening the question of whether the voice carries more deception signal than the face, a theme that overlaps with speech emotion recognition accuracy research. We aggregated data from the Journal of Forensic Sciences, Columbia University’s deception speech corpus, PLOS One, the European Journal of Psychology Applied to Legal Context, the U.S. Department of Labor, the EU AI Act and UK government records.
TL;DR
- Voice stress analyzers caught about 15% of lies about drug use, with about 50% overall accuracy, in a jail field test of 319 arrestees (NIJ, Voice Stress Analysis field test 2008).
- NITV-provided CVSA examiners in a lab test flagged truthful speech as deceptive 61% to 70% of the time (Journal of Forensic Sciences, Evaluation of the NITV CVSA 2008).
- The research offers “little or no scientific basis” for the computer voice stress analyzer (National Research Council, The Polygraph and Lie Detection 2003).
- People average 54% correct lie-truth judgments across 206 documents (Personality and Social Psychology Review, Accuracy of Deception Judgments 2006).
- Human interviewers scored 56.75% on the Columbia deception corpus, while a speech classifier reached an F1 of 72.77 (Columbia University, Interspeech 2018).
- Listening to audio only gave 61.7% accuracy vs 35% for watching video with sound, in a test with 120 participants (European Journal of Psychology Applied to Legal Context 2025).
- A 2026 test with 246 participants found no significant difference between audio-only, transcript and audio-visual judgments (PLOS One 2026).
- Higher pitch while lying is a small effect, d = 0.21, rising to d = 0.59 for motivated liars (Psychological Bulletin, Cues to Deception 2003).
- Only 1 of 12 UK councils that used voice risk analysis in benefit reviews reported it as a success (DWP evaluation, reported by The Register 2010).
- U.S. law treats voice stress analyzers as lie detectors (U.S. Department of Labor, EPPA Fact Sheet #36), and the EU classifies AI lie-detection tools for police as high-risk (Regulation (EU) 2024/1689).
1. Field Tests: Voice Stress Analyzers Against Verified Ground Truth
Field tests are the only place where voice stress analysis meets a real outcome it cannot influence, such as a urine test. In the NIJ-funded Oklahoma County jail study (NIJ Journal field test summary, final report), arrestees answered questions about drug use and their answers were checked against urinalysis. LVA flagged 21% of the deceptive answers and CVSA flagged 8%, while both programs labeled most truthful answers correctly. A tool that says “truthful” almost every time will look accurate on honest people and miss nearly every liar. The final report also found that scoring was unreliable: correlations between novice and expert scorers ranged from 0.11 to 0.52, depending on the drug.
The most interesting number is not an accuracy rate at all. Arrestees who believed a voice device was monitoring them lied far less, a “bogus pipeline” effect (people tell more truth when they believe a machine can catch lies). The device may deter lying without detecting it.
| Metric | Value | Source |
|---|---|---|
| Overall accuracy of CVSA and LVA in detecting deception about recent drug use (319 male jail arrestees, Oklahoma County) | About 50% | NIJ, Voice Stress Analysis field test (NIJ Journal 259) 2008 |
| Deceptive answers correctly flagged as deceptive by LVA | 21% | NIJ, Voice Stress Analysis field test 2008 |
| Deceptive answers correctly flagged as deceptive by CVSA | 8% | NIJ, Voice Stress Analysis field test 2008 |
| Truthful answers correctly identified as truthful, average across five drugs (LVA 95%, CVSA 90%) | 91.5% | NIJ, Voice Stress Analysis field test 2008 |
| Arrestees who lied about recent drug use: voice stress study vs earlier ADAM drug-monitoring study in the same city | 14% vs 40% | NIJ, Voice Stress Analysis field test 2008 |
| Military interrogation-training (SERE) field recordings: deceptive yes/no answers flagged by the CVSA maker’s own examiners vs truthful answers wrongly flagged | 19% vs 45% | Journal of Forensic Sciences, Evaluation of the NITV CVSA 2008 |
| Same field recordings scored by experimental phoneticians: deceptive answers flagged vs truthful answers wrongly flagged | 38% vs 49% | Journal of Forensic Sciences, Evaluation of the NITV CVSA 2008 |
Context: the earlier ADAM survey took about 20 minutes and the voice stress survey about 5 minutes, so the 14% vs 40% gap is not a controlled experiment. Most recent available data: NIJ, 2007 report and 2008 article. We found no comparable independent field test published since.
2. Laboratory Tests of CVSA and LVA
Laboratory studies give the devices their best possible conditions: clear recordings, known truth and induced stress. They still failed. In the University of Florida evaluations (CVSA study, LVA study), 48 speakers produced 300 samples under high-jeopardy lies and stress. The CVSA maker’s own examiners flagged truthful speech as deceptive 61% to 70% of the time, about as often as they caught real lies.
The sensitivity measure tells the same story. d-prime is a signal-detection score where 0 means no ability to separate two categories; the authors set 1 as the minimum acceptable value, and every team scored far below it. The National Research Council review adds that the one perfect result it found came from a vendor-affiliated article that never described how it scored charts or established the truth.
| Metric | Value | Source |
|---|---|---|
| CVSA lab test (48 speakers, 300 samples): hit rate of NITV-provided examiners | 61% to 65% | Journal of Forensic Sciences, Evaluation of the NITV CVSA 2008 |
| CVSA lab test: false-positive rate (truthful or low-stress speech flagged) of NITV-provided examiners | 61% to 70% | Journal of Forensic Sciences, Evaluation of the NITV CVSA 2008 |
| CVSA lab test: d-prime sensitivity of NITV-provided examiners (minimum acceptable value set at 1) | -0.32 to 0.13 | Journal of Forensic Sciences, Evaluation of the NITV CVSA 2008 |
| LVA lab test: hit rates across conditions and examiners | 42% to 56% | Journal of Forensic Sciences, Evaluating Layered Voice Analysis 2009 |
| LVA lab test: false-positive rates, often above the hit rates | 40% to 65% | Journal of Forensic Sciences, Evaluating Layered Voice Analysis 2009 |
| Agreement between CVSA and polygraph in a vendor-affiliated study of 54 probationers (scoring and truth criteria not described) | 100% (vendor-affiliated claim, judged unusable) | National Research Council, The Polygraph and Lie Detection 2003 |
| CVSA accuracy vs polygraph in 1990s Department of Defense Polygraph Institute lab comparisons | Never significantly superior | National Research Council, The Polygraph and Lie Detection 2003 |
| Results of a 1996 field test of the analyzer | Close to chance levels | National Research Council, The Polygraph and Lie Detection 2003 |
Context: the same NRC chapter reports that agreement between independent analysts reading the same voice stress charts “is generally low”, and that a 2000 Defense Department study found analyzer scores did not reflect acute stress measured by traditional methods.
3. How Accurately Humans Judge Lies (Meta-Analytic Evidence)
If machines cannot hear lies, the obvious fallback is a trained ear. The evidence says ears are only marginally better than eyes. The Bond and DePaulo meta-analysis pooled 206 documents and 24,483 judges. People average 54% correct, and they call truths correctly far more often than lies, because most judges lean toward believing what they hear (55.23% of messages were judged truthful, weighted). Experts scored 54.51% vs 53.31% for non-experts between studies, and the gap disappeared after statistical adjustment.
The medium comparison is the relevant one for voice. Silent video is the worst condition, and adding sound recovers a few points. Hearing a liar helps a little; it does not turn listeners into lie detectors.
| Metric | Value | Source |
|---|---|---|
| Average correct lie-truth judgments (206 documents, 24,483 judges) | 54% | Personality and Social Psychology Review, Accuracy of Deception Judgments 2006 |
| Truthful messages correctly judged as truthful | 61% | Personality and Social Psychology Review, Accuracy of Deception Judgments 2006 |
| Deceptive messages correctly judged as deceptive | 47% | Personality and Social Psychology Review, Accuracy of Deception Judgments 2006 |
| Correct lie-truth judgments with audio-only messages | 53.01% | Personality and Social Psychology Review, Accuracy of Deception Judgments 2006 |
| Correct lie-truth judgments with silent video-only messages | 50.52% | Personality and Social Psychology Review, Accuracy of Deception Judgments 2006 |
| Correct lie-truth judgments with audiovisual messages | 54.03% | Personality and Social Psychology Review, Accuracy of Deception Judgments 2006 |
Most recent available data: Personality and Social Psychology Review, 2006. It remains the largest synthesis of human lie-detection accuracy; section 4 covers the newer listening experiments.
4. Listening Experiments and Speech Classifiers, 2018 to 2026
Newer studies test the voice directly, and they disagree. A 2025 University of Portsmouth experiment (European Journal of Psychology Applied to Legal Context) gave 120 participants either the audio or the full video of a mock-suspect interview. Listeners were right 61.7% of the time, while those who watched and listened were right only 35% of the time, below chance. The study treated watching plus listening as the high cognitive load condition: seeing the suspect added information that, in this test, hurt more than it helped. A larger 2026 test (PLOS One) found no modality effect at all, with liars mostly judged truthful in every condition.
Machines trained on interview speech do better than people on the same recordings. On the Columbia X-Cultural Deception corpus (NAACL 2018, Interspeech 2018), 340 speakers gave truthful and deceptive answers in interview dialogues. A classifier using acoustic-prosodic features (pitch, loudness and timing) matched one built on word choice. These are lab lies on one corpus, and they fall well short of the reliability expected of voice systems with a hard ground truth, such as those tracked in speaker verification benchmarks.
| Metric | Value | Source |
|---|---|---|
| Human interviewers’ lie-truth accuracy on the Columbia X-Cultural Deception corpus (340 speakers; chance = 50%) | 56.75% | Columbia University, NAACL 2018 (Levitan et al.) |
| Human interviewers’ accuracy on deceptive answers only | 47.93% | Columbia University, NAACL 2018 (Levitan et al.) |
| Best classifier using word-use (lexical) features, F1 score | 72.74 (accuracy 72.78%) | Columbia University, NAACL 2018 (Levitan et al.) |
| Best classifier using acoustic-prosodic features plus speaker gender and native language, F1 score | 72.77 | Columbia University, Interspeech 2018 (Levitan et al.) |
| Overall lie-truth accuracy: listening to audio only vs watching video with sound (120 participants) | 61.7% vs 35% | European Journal of Psychology Applied to Legal Context 2025 (Giorgianni et al.) |
| Truth detection: audio only vs video with sound | 53.33% vs 20.00% (chi-square(1, N = 60) = 7.18, p = .007) | European Journal of Psychology Applied to Legal Context 2025 (Giorgianni et al.) |
| Audio-only condition: truthful statements judged correctly vs deceptive statements judged correctly (N = 246 in total) | 60.71% vs 39.29% | PLOS One 2026 (Bogaard et al.) |
| Effect of modality (audio only, transcript, audio-visual) on lie-truth accuracy | F(2, 243) = 0.45, p = .637 (not significant) | PLOS One 2026 (Bogaard et al.) |
Context: in the 2025 study the lie-detection difference between conditions (70.00% vs 50.00%) was not significant (p = .114); the significant gain came from recognizing truthful suspects.
5. Acoustic and Vocal Cues of Lying
Voice stress devices assume lying leaves a strong, universal acoustic trace. The meta-analytic record shows a few faint traces. The Psychological Bulletin meta-analysis summarized 1,338 effect sizes for 158 cues from 120 independent samples. Pitch rose only slightly in liars overall (d = 0.21), but jumped to d = 0.59 when liars were motivated to succeed. Cohen’s d is a standardized effect size, and values this size are conventionally labeled small: most vocal cues overlap heavily between honest and dishonest speakers.
Single-language studies show how unstable the cues are. In 19 native Italian speakers, mean pitch was 160.88 Hz when lying vs 160.67 Hz when telling the truth (p = 0.927), but speech slowed (Frontiers in Psychology). In 36 Mandarin speakers answering 14 biographical questions, a 2024 study of 20 acoustic and 14 electroglottographic features (electroglottography measures vocal-fold contact) linked deception to higher mean pitch and intensity and lower jitter, shimmer and harmonics-to-noise ratio (Journal of Speech, Language, and Hearing Research). The Columbia corpus showed higher maximum pitch and maximum intensity in deceptive answers. Cues like these are the same raw material used in paralinguistic voice analysis for burnout prediction, and they share the same limit: they describe groups, not individuals.
| Metric | Value | Source |
|---|---|---|
| Higher pitch when lying, all studies (12 estimates) | d = 0.21 (95% CI 0.08 to 0.34) | Psychological Bulletin, Cues to Deception 2003 |
| Vocal tension when lying (10 estimates) | d = 0.26 (95% CI 0.13 to 0.39) | Psychological Bulletin, Cues to Deception 2003 |
| Liars sounding vocally uncertain (10 estimates) | d = 0.30 | Psychological Bulletin, Cues to Deception 2003 |
| Talking time of liars vs truth-tellers (4 estimates; negative means liars talk less) | d = -0.35 | Psychological Bulletin, Cues to Deception 2003 |
| Filled pauses (um, uh) in lies vs truths (16 estimates) | d = 0.00 | Psychological Bulletin, Cues to Deception 2003 |
| Pitch effect when liars were motivated to succeed (6 estimates) | d = 0.59 (95% CI 0.31 to 0.88) | Psychological Bulletin, Cues to Deception 2003 |
| Pitch and response latency, meta-analysis of 41 studies covering nine speech cues (both higher when lying) | d = 0.268 (pitch), d = 0.177 (latency) | Applied Cognitive Psychology meta-analysis 2006 (Sporer and Schwandt), as reported in Frontiers in Psychology 2012 |
| Speech rate when lying vs telling the truth, 19 native Italian speakers | 1.95 vs 2.10 words per second (d = 0.447, p = 0.025) | Frontiers in Psychology, Markers of Deception in Italian Speech 2012 |
Context: response latency is the clearest conflict between the two meta-analyses. The 2006 synthesis reports a reliable increase, while the 2003 synthesis lists a near-zero, non-significant effect across 32 estimates.
6. Use by Police Agencies, Governments and Insurers
Adoption has never tracked the evidence. The adoption figures come almost entirely from the manufacturer, and the claimed totals have grown since independent tests found chance-level accuracy. A San Francisco Chronicle investigation published on June 4, 2024 identified 13 other California law enforcement agencies using the CVSA, three of which had also used it in criminal investigations (San Francisco Chronicle).
Government trials outside policing ended differently. The UK Department for Work and Pensions ran “voice risk analysis” trials on benefit claims by phone; most councils that reported results did not call it a success (The Register, November 10, 2010). UK insurers such as esure were already piloting voice risk analysis on claims lines in 2003, and newer vendors now pitch voice screening from application to claim; for example, Clearspeed states on its own website that a Department of Defense validation evaluated its technology at above 97% accuracy, a vendor’s own claim for which we found no published independent study. Insurers that already run speech analytics for sentiment detection face the same question: what is actually being measured?
| Metric | Value | Source |
|---|---|---|
| U.S. law enforcement agencies using CVSA, according to its manufacturer (2008) | More than 1,400 | NITV figure reported in NIJ Journal 2008 (vendor’s own claim) |
| Agencies using CVSA, according to the manufacturer’s website (2026) | Over 3,000 U.S. law enforcement agencies; over 3,600 worldwide | NITV website 2026 (vendor’s own claim) |
| Other California law enforcement agencies identified as using CVSA | 13 | San Francisco Chronicle investigation, June 4, 2024 |
| Of those agencies, number that also used CVSA in criminal investigations | 3 | San Francisco Chronicle investigation, June 4, 2024 |
| Cost of the UK DWP voice risk analysis trials on benefit claims | £2.16m | DWP voice risk analysis evaluation, reported by The Register, November 10, 2010 |
| Local authorities in the latest batch of DWP trials, and how many reported results | 24 (23 reported) | DWP voice risk analysis evaluation, reported by The Register, November 10, 2010 |
| Councils using voice risk analysis in benefit reviews that reported it as a success | 1 of 12 | DWP voice risk analysis evaluation, reported by The Register, November 10, 2010 |
| Councils using it on calls from new claimants that reported it as a success | 3 of 9 | DWP voice risk analysis evaluation, reported by The Register, November 10, 2010 |
Context: the manufacturer’s own website is internally inconsistent, citing both “over 3,000” law enforcement agencies and “over 2,900 U.S. organizations” in different places.
7. Scientific and Legal Assessments
The scientific verdict has been stable for two decades, and the law has slowly caught up. The National Research Council concluded in 2003 that voice stress research offers “little or no scientific basis” for the computer voice stress analyzer. In the United States the strongest restriction is on employers: the Employee Polygraph Protection Act names the voice stress analyzer as a lie detector, which bars most private employers from using it. Federal, state and local governments are excluded, so police agencies fall outside that ban.
The European Union regulates the newer, AI-based generation. The AI Act lists AI systems used “as polygraphs or similar tools” by law enforcement and by migration and border authorities as high-risk, and its recitals call for banning AI that reads emotions at work or in schools. The UK government’s own conclusion came in a 2011 freedom of information response: it no longer used the technology.
| Metric | Value | Source |
|---|---|---|
| Scientific basis for using the computer voice stress analyzer | ”Little or no scientific basis” | National Research Council, The Polygraph and Lie Detection 2003 |
| Devices covered by the U.S. legal definition of a lie detector | Polygraph, deceptograph, voice stress analyzer, psychological stress evaluator or similar device | U.S. Department of Labor, EPPA Fact Sheet #36 |
| Lie detector tests private employers may still use, and in how many limited situations | Polygraph only (no other lie detector), in 3 limited situations | U.S. Department of Labor, EPPA Fact Sheet #36 |
| Deadline for employees to file a lawsuit for an EPPA violation | 3 years | U.S. Department of Labor, EPPA Fact Sheet #36 |
| EU AI Act, Annex III point 6(b): AI used by law enforcement as polygraphs or similar tools | Classified high-risk | Regulation (EU) 2024/1689 (AI Act) |
| EU AI Act, Annex III point 7(a): AI used by migration, asylum and border authorities as polygraphs or similar tools | Classified high-risk | Regulation (EU) 2024/1689 (AI Act) |
| EU AI Act, Recital 44: AI systems that detect people’s emotional state at work or in education | To be prohibited | Regulation (EU) 2024/1689 (AI Act) |
| UK DWP conclusion on voice risk analysis for benefit claims | Could not conclude it worked effectively and consistently; not in use | DWP Freedom of Information response 3039/2011 |
Context: the EU AI Act text was adopted on June 13, 2024. The EPPA was signed on June 27, 1988 and took effect on December 27, 1988.
Summary: Voice Stress Analysis and Vocal Lie Detection by the Numbers
| Metric | Value | Source |
|---|---|---|
| Deceptive answers about drug use flagged in a jail field test: LVA vs CVSA | 21% vs 8% | NIJ, Voice Stress Analysis field test 2008 |
| Military field recordings: deceptive answers flagged vs truthful answers wrongly flagged by CVSA maker’s examiners | 19% vs 45% | Journal of Forensic Sciences, Evaluation of the NITV CVSA 2008 |
| Scientific basis for the computer voice stress analyzer | ”Little or no scientific basis” | National Research Council, The Polygraph and Lie Detection 2003 |
| Average human lie-truth accuracy (24,483 judges) | 54% | Personality and Social Psychology Review, Accuracy of Deception Judgments 2006 |
| Human interviewers on the Columbia deception corpus | 56.75% | Columbia University, NAACL 2018 (Levitan et al.) |
| Best acoustic-prosodic classifier on the same corpus, F1 score | 72.77 | Columbia University, Interspeech 2018 (Levitan et al.) |
| Accuracy listening to audio only vs watching video with sound | 61.7% vs 35% | European Journal of Psychology Applied to Legal Context 2025 (Giorgianni et al.) |
| Effect of audio-only vs transcript vs audio-visual format on accuracy | F(2, 243) = 0.45, p = .637 | PLOS One 2026 (Bogaard et al.) |
| Audio-only judgments: truthful statements judged correctly vs deceptive statements judged correctly | 60.71% vs 39.29% | PLOS One 2026 (Bogaard et al.) |
| Higher pitch when liars are motivated to succeed | d = 0.59 | Psychological Bulletin, Cues to Deception 2003 |
| Pitch effect in a 41-study meta-analysis of speech cues | d = 0.268 | Applied Cognitive Psychology meta-analysis 2006 (Sporer and Schwandt), as reported in Frontiers in Psychology 2012 |
| Speech rate lying vs truth-telling, 19 Italian speakers | 1.95 vs 2.10 words per second | Frontiers in Psychology, Markers of Deception in Italian Speech 2012 |
| Other California agencies identified using CVSA | 13 | San Francisco Chronicle investigation, June 4, 2024 |
| Cost of UK voice risk analysis benefit trials | £2.16m | DWP voice risk analysis evaluation, reported by The Register, November 10, 2010 |
| Councils reporting success with voice risk analysis in benefit reviews | 1 of 12 | DWP voice risk analysis evaluation, reported by The Register, November 10, 2010 |
| Voice stress analyzer in the U.S. legal definition of a lie detector | Included | U.S. Department of Labor, EPPA Fact Sheet #36 |
| AI used by police as polygraphs or similar tools in the EU | High-risk (Annex III point 6(b)) | Regulation (EU) 2024/1689 (AI Act) |
Methodology and Sources
- National Institute of Justice: Voice Stress Analysis: Only 15 Percent of Lies About Drug Use Detected in Field Test (NIJ Journal 259, March 2008) and the final report, Assessing the Validity of Voice Stress Analysis Tools in a Jail Setting (June 2007).
- National Research Council: The Polygraph and Lie Detection, Chapter 6, Alternative Techniques and Technologies (2003).
- Journal of Forensic Sciences (University of Florida): Evaluation of the NITV CVSA (2008) and Stress and Deception in Speech: Evaluating Layered Voice Analysis (2009).
- Personality and Social Psychology Review: Accuracy of Deception Judgments (Bond and DePaulo, 2006).
- Psychological Bulletin: Cues to Deception (2003).
- Frontiers in Psychology: Markers of Deception in Italian Speech (Spence, Villar and Arciuli, 2012), also the source we used for the Applied Cognitive Psychology meta-analysis figures (Sporer and Schwandt, 2006).
- Columbia University speech lab: Linguistic Cues to Deception and Perceived Deception in Interview Dialogues (NAACL 2018) and Acoustic-Prosodic Indicators of Deception and Trust in Interview Dialogues (Interspeech 2018).
- European Journal of Psychology Applied to Legal Context: The Effect of the Interviewer’s Cognitive Load on the Quality of the Forensic Interview (2025).
- PLOS One: I Told You So! Verbal cue beliefs are associated with truth detection, but not lie detection (January 23, 2026).
- Journal of Speech, Language, and Hearing Research: Uncovering Gender-Specific and Cross-Gender Features in Mandarin Deception: An Acoustic and Electroglottographic Approach (2024).
- U.S. Department of Labor, Wage and Hour Division: Fact Sheet #36, Employee Polygraph Protection Act.
- European Union: Regulation (EU) 2024/1689, the Artificial Intelligence Act.
- UK Department for Work and Pensions: Freedom of Information response 3039/2011; trial figures as reported by The Register (November 10, 2010), because the original DWP evaluation file is no longer reachable online.
- San Francisco Chronicle: investigation into CVSA use by California law enforcement (June 4, 2024).
- Vendor claims (labeled as such, not linked): NITV website (agency counts) and Clearspeed website (validation accuracy), both read November 2026.
- Data watch: the core independent accuracy data on commercial voice stress analyzers is 17 to 23 years old (2003 to 2009), and we found no newer independent field validation; treat 2026 vendor accuracy and adoption figures as unverified marketing. The two speech meta-analyses in section 5 conflict on response latency. Small-corpus machine-learning studies report much higher numbers than the Columbia work; one 2025 book chapter on 121 courtroom video clips gives its audio model’s accuracy as 93.57% in one table and 95.57% in the text, so we left such results out of the headline tables. The 2025 and 2026 listening studies point in opposite directions on whether audio-only judgments beat audio-visual ones, so a new meta-analysis is needed before drawing a conclusion.
Last updated: November 8, 2026. We update this roundup quarterly, and the next refresh is expected when the next independent field validation of a commercial voice stress analyzer is published in a peer-reviewed journal.