NIJ Voice Stress Study: What the 50% Finding Means

Research evidence at a glance

These sources address different questions. Publication date alone does not determine relevance: the system, participants, task, and outcome matter. The 2008 NIJ Journal article summarizes the 2007 report; it is not a separate field experiment.

Read the finding and its scope

The National Institute of Justice reported a sobering result from a field test of voice stress analysis. Two programs, CVSA and LVA, performed poorly at identifying deceptive answers about recent drug use among arrestees. This is a finding to take seriously when evaluating voice stress equipment.

The study did not evaluate every voice stress system, and it did not test VIPRE’s VASIS or AISIS. That distinction limits a claim about those specific products in either direction. The NIJ finding does not directly measure them, and it does not validate them.

What happened in the field test?

Researchers interviewed 319 recently arrested men at an Oklahoma County jail about recent use of several drugs. Urine tests provided an external check on the answers. Researchers collected voice stress results from CVSA and LVA and compared them with that check.

The NIJ reports that the test formats were approved by officials from both companies and that the data were interpreted by researchers and certified company examiners.

What the numbers mean

Across the drug questions, the programs identified an average of 15 percent of deceptive responses as deceptive. The programs classified truthful responses as truthful more often. A system can appear reassuring on one figure and still miss most deceptive answers.

The NIJ’s approximate 50 percent result summarized the programs’ combined ability to separate the two categories in that setting. It was not a claim that each individual answer had a 50 percent chance of being correct.

What the study does and does not establish

The study is evidence about the systems, versions, population, questions, and procedure studied. It identifies CVSA and LVA, not VASIS or AISIS. Yet saying a device was not studied does not answer the next question: what independent, product-specific evidence supports its performance?

Agencies should request that evidence, including both missed deceptive answers and false alarms, before relying on any accuracy claim.

A practical reading for investigators

Treat a voice stress indication as a prompt to investigate, not a verdict. Record the question and the context, consider ordinary sources of stress, and check relevant facts independently. Do not rely on a vendor percentage without a clear method and a study of the product being offered.

Separate product documentation from validation

AISIS provides microphone-only AI-assisted voice stress analysis and audio review. AISIS-K adds camera capture and kinesic video recall. VASIS provides an examiner-led voice stress workflow. These descriptions identify product capabilities; they are not independent findings about stress classification or deception detection.

Before relying on a performance claim, request the full report and check the intended use, participants, independent reference, blinded assessment, false alarms, missed cases, and inconclusive outcomes. An AI label, a demonstration, certification, or an operational reference does not establish accuracy.

• Product documentation: what the system records, displays, stores, and exports
• Practitioner experience: how trained examiners use the workflow, with permission and context
• Internal evaluation: who ran it, funding, method, sample, and all outcomes disclosed
• Independent validation: the exact product and version tested against a defined external reference

Primary sources

Read the NIJ final report, its plain-language Journal summary, and the National Academies review for the source context behind this guide.