A Novel Approach for Enhancing Audio Modality in Multimodal Deception Detection Environment Using Liquid Neural Networks
Ahmed T. Elbattawy, Khaled Mahmoud Badran, Mohamed K. Elhadad · 2025
Deception detection is a critical research area with applications in security, forensics, and psychology. Traditional multimodal deception detection models incorporate audio and visual modalities, but audio-based deception detection often remains underexplored. This study introduces Liquid Neural Networks (LNNs) to enhance the audio modality, improving deception detection accuracy. Unlike conventional Recurrent Neural Networks (RNNs), LNNs dynamically adjust their internal weights based on input strength, providing superior adaptability to speech variations. Our experiments on the Dolos dataset and a real-life courtroom deception dataset demonstrate that LNNs achieve an accuracy of 65.33% on the Dolos dataset and 87.5% on the courtroom dataset, outperforming traditional machine learning models. Compared to the state-of-the-art models, LNNs achieve a 5.04% improvement over the best prior audio-based method on the Dolos dataset and an 8.5% increase over previous benchmarks on the courtroom dataset. This study highlights the potential of LNNs in deception detection and their applicability to real-world forensic investigations.