The research we build on, with sources.
Veriteus builds on published research into deceptive speech. Here is what that research shows, where its limits are and how we use it.
People get it right 54% of the time.
A meta-analysis of 206 documents and 24,483 judges found that people told lies from truth with 54% accuracy, barely above chance. Experienced judges did no better than beginners. Decisions in claims, hiring and investigations still rest on that judgment.
Speech carries signals people miss.
Since 2005, Prof. Julia Hirschberg and colleagues at Columbia University have built recorded interview corpora in which speakers lie or answer honestly for a reward. Models trained on acoustic, prosodic and lexical features told the two apart better than chance, and better than human listeners on the same recordings.
What these studies don’t show
- They were run in a lab, with small rewards and no real consequences.
- Speakers were native speakers of American English and Mandarin Chinese, speaking English.
- The way people lie differs by native language and gender, so results don’t carry over to other languages without new data.
These are research results. They are not measurements of Veriteus.
Selected papers
How this differs from voice-stress analyzers.
Voice-stress analyzers claim to read deception from stress in the voice. In a field study funded by the U.S. National Institute of Justice, researchers tested two of them, CVSA and LVA, on people held in a county jail and compared what they said about recent drug use with urine tests. Neither tool detected deception better than chance.
Forensic specialists we interviewed while building Veriteus had stopped using a voice analyzer for the same reason. In their words, it was fifty-fifty.
What Veriteus does differently
- It reads many features of speech, calibrated to each speaker.
- It checks what is said against the file, so a flag can rest on a mismatch in content.
- Every flag comes with a follow-up question. The answer shows whether the concern holds.
- When we share our own results, we share the conditions they were measured under.
What Veriteus has measured so far.
Our own numbers, with the conditions they were measured under. They are separate from the research above.
Over 70%
Accuracy in a controlled setting, with longer conversations and each speaker recorded on a separate channel.
4 in 5
When Veriteus flags a moment, it’s right about four times in five. It doesn’t catch every one.
Over 80%
Our target when voice and video are combined. Video analysis is on our roadmap.
Output is input for the person asking, never a verdict.
Output for a person who asks, never a ruling.
Veriteus combines voice signals with a check of what is said against the file. The result is a marked passage and a suggested question. A person reads it in context and decides what to do.
Results depend on language, call length and the documents available. That is why every engagement starts with a pilot on your own closed cases.
Who advises us.
- Prof. Julia Hirschberg, Columbia University. Lead scientific advisor.
- Dr. Sarah Ita Levitan, Hunter College.
- Doc. Jindřich Matoušek, University of West Bohemia. Speech research.
- Marek Hrúz, Ph.D., University of West Bohemia. Video research.
Questions about our method?
Send them to Tomáš Šebesta, our CSO. We’ll answer them, including where the method has its limits.