The Sound of Deception: Voice and Language Markers in Emotionally Charged Lies

Detecting deception in spoken language is a challenge in law enforcement, security, psychology, and everyday interactions. While studies have identified various acoustic and linguistic traits that may indicate deception, human judgments continue to be inconsistent and often unreliable. Consequently, there is increasing interest in developing objective methods to better understand and identify deceptive communication.

This study examined the acoustic and linguistic markers of deception in spoken Croatian by analyzing narrative speech, intending to identify patterns that distinguish deceptive from truthful accounts across various emotional contexts. Participants were asked to provide three types of stories: a fabricated account, a truthful story with negative emotional content, and a truthful story with positive emotional content. All narratives were audio-recorded during the session and subsequently transcribed.

Acoustic features such as fundamental frequency (pitch), intensity (loudness), speech rate, and pause frequency were extracted using Praat and SIVE software to explore their relevance as indicators of cognitive load and stress during speech production. Linguistic analysis focused on syntactic complexity, lexical choices, and the usage of personal pronouns and affective language, with linguistic features automatically quantified from Croatian transcripts using the xAI API and the Grok-3 large language model.

Differences in acoustic and linguistic variables across the three conditions were statistically examined to identify patterns associated with deception. To further explore whether these features could distinguish deceptive from truthful speech, several machine learning classification algorithms were applied. Models were trained separately on acoustic features, linguistic features, and their combination to assess the potential of a multimodal approach.

The presentation will demonstrate how emotional states and cognitive load influence acoustic and linguistic markers of deception and will discuss the implications of these findings for developing voice-based lie detection tools. Examples of the most informative speech features and their patterns across conditions will be highlighted to illustrate key trends.