Explainable Enhanced Recurrent Neural Network for lie detection using voice stress analysis

Abstract

Lie detection is a crucial aspect of human interactions that affects everyone in their daily lives. Individuals often rely on various cues, such as verbal and nonverbal communication, particularly facial expressions, to determine if someone is truthful. While automated lie detection systems can assist in identifying these cues, current approaches are limited due to a lack of suitable datasets for testing their performance in real-world scenarios. Despite ongoing research efforts to develop effective and reliable lie detection methods, this remains a work in progress. The polygraph, voice stress analysis, and pupil dilation analysis are some of the methods currently used for this task. In this study, we propose a new detection algorithm based on an Enhanced Recurrent Neural Network (ERNN) with Explainable AI capabilities. The ERNN, based on long short-term memory (LSTM) architecture, was optimized using fuzzy logic to determine the hyperparameters. The LSTM model was then created and trained using a dataset of audio recordings from interviews with a randomly selected group. The proposed ERNN achieved an accuracy of 97.3%, which is statistically significant for the problem of voice stress analysis. These results suggest that it is possible to detect patterns in the voices of individuals experiencing stress in an explainable manner.