Voice Stress Analysis: A Technology That Does Not Work

There is a device that police forces, insurers and employers have bought in quantity for decades. You speak into it. Software analyzes your voice for involuntary microtremors said to be suppressed under stress. It reports whether you are lying.

It does not work. This is not a contested question.

The National Institute of Justice funded a field study at a county jail, testing computer voice stress analysis on arrestees who were asked about recent drug use, with independent urinalysis providing the ground truth. The two commercial systems tested detected roughly fifteen percent of the deceptive responses. A coin flip would do better. A related NIJ-funded laboratory evaluation reached comparable conclusions, and the broader review literature has consistently failed to find validity for the underlying premise. The microtremor itself — a purported eight-to-twelve hertz oscillation in the muscles that would be damped under stress — has never been demonstrated in the laryngeal musculature in the way the marketing describes.

So why do police departments keep buying them, and why do the vendors keep producing testimonials from satisfied investigators?

Because the device works. It just does not work by measuring anything.

The mechanism is the bogus pipeline, an experimental technique from social psychology in which subjects are connected to an impressive apparatus they are told can detect their true attitudes, and consequently report those attitudes more honestly. Belief in the instrument does the work. A suspect who is told that a machine has just detected deception in his voice, and who believes it, is in a materially different position: denial now appears futile, the interrogator's confidence appears evidence-based, and confession becomes the rational move. Confessions follow. The investigator observes that the machine identified a liar. The machine identified nothing.

This is worth sitting with, because it generalizes far beyond one product category.

The polygraph, examined in chapters 114 and 115, has a substantially better-founded physiological basis and a substantially worse reputation than its accuracy warrants in some applications — but a large part of its operational value has always been the same effect, and the National Academies' 2003 review said as much. So-called brain fingerprinting, in chapter 117, has the same structure: a real signal, an unvalidated inferential leap, and enormous persuasive force in a room.

And the failure mode is not merely that the innocent are inconvenienced. An instrument that produces confessions through belief will produce false confessions at the same rate it produces true ones, because the pressure it applies is indifferent to guilt. Chapters 92, 96 and 97 are what that looks like in practice.

The general principle to take away is a question to ask of any instrument offered as evidence of a mental state: what is its measured accuracy against an independent ground truth, established by someone with no commercial interest, and published? For voice stress analysis the answer exists, and it is fifteen percent.

Any technology whose vendors offer case testimonials instead of that number is telling you which kind of instrument it is.