The Lie Detector: Combining Fact Verification with Within-Speaker Speech Analysis for AI-Assisted Lie Detection

Existing lie-detection tools each answer only part of the question of whether a person’s words are true or false. The polygraph requires sensors and the subject’s physical presence; commercial voice stress analysers return scores from closed algorithms; fact-checking services establish whether a claim is false but do not consider how it was made. This paper presents the Lie Detector, an AI-assisted method that issues one of three verdicts — Truth, Lie or Not established — for every statement in an interview, speech or written text. The method combines two lines of evidence. Fact verification checks each checkable claim against official sources and attaches the evidence. Within-speaker speech analysis compares each answer with the same speaker’s truthful accounts of comparable events, so that no population norm is required. A Lie verdict requires both a documented falsehood and a stable change in speech in the same passage; a false claim without such a change, a disputed claim or an uncheckable claim is reported as Not established. Every analysis passes an automatic double check, and every Lie verdict is approved and signed by a qualified expert. The method works from recordings or texts in Russian, English and Chinese (voice in Russian and English). The paper describes the method and its decision rule, compares it with existing approaches, demonstrates its operation on synthetic material and sets out an evaluation design.