An AI voice detector analyzes speech audio to estimate whether it was produced by a human or generated by a voice-synthesis model. Demand for the category has followed the collapse in cloning difficulty: modern systems can produce a convincing clone of a specific person from seconds of sample audio, which turned vishing from a generic scam into targeted executive impersonation.

What detectors actually look for

  • Spectral artifacts. Synthesis models reconstruct audio from compressed representations, leaving statistical fingerprints in frequency bands humans don't consciously perceive.
  • Prosody and breathing. Micro-timing of breaths, hesitations, and pitch drift are hard to fake consistently across a long conversation.
  • Channel consistency. Real calls carry coherent room acoustics and microphone characteristics; spliced or generated audio often doesn't.
  • Liveness signals. Interactive tests — unexpected questions, requests to vary speech — stress generators in ways static playback doesn't. This is the voice equivalent of liveness detection in identity verification.

Reading accuracy claims skeptically

Vendor accuracy numbers are usually measured on benchmark datasets, against known generators, with clean audio. Three things degrade real-world performance:

  1. Compression. Phone codecs and conferencing platforms strip exactly the high-frequency detail many detectors rely on.
  2. Novel generators. Detection is an arms race; a detector trained against last year's models has never seen this year's.
  3. Base rates. If one call in ten thousand is an attack, even a 99%-accurate detector produces far more false alarms than true detections. Alert fatigue follows.

The right architecture: detection as a trigger

Because no detector is perfect, the mature deployment pattern treats the risk score as a trigger for a deterministic control rather than a verdict. A suspicious score on a call that requests a payment should automatically require out-of-band verification — a callback to a registered number — before anything moves. The detector narrows attention; the protocol provides the guarantee. This is the model VerifyTheCall implements: scoring where audio access exists (Zoom, Teams, SIP), enforced callback verification everywhere, including channels where no detector can listen.

Protocol tipEvaluate any voice detector by asking one question: what happens operationally when the score is 70%? If the answer is "someone decides," you have an alert, not a control.