Real-Time Lie-Speech Determination Using Voice-Stress Technology

Fadi Al-Dhaher; Duraid  Y. Mohammed; Mohammed  Khalaf; Khamis  Al-Karawi; Mohammad  Sarfraz; Muhammad Mazin  Al Maathidi

doi:10.52866/ijcsm.2024.05.02.008

Authors

Fadi Al-Dhaher Computer Engineering Department, College of Engineering, Al-Iraqia University,
Duraid Y. Mohammed Computer Engineering Department, College of Engineering, Al-Iraqia University, Baghdad, Iraq https://orcid.org/0000-0002-9586-1983
Mohammed Khalaf General Directorate of Education Anbar, Ramadi 31001, Iraq / Computer Science Department, Al-Maarif University College, Ramadi 31001, Iraq https://orcid.org/0000-0003-0559-455X
Khamis Al-Karawi University of Diyala, Iraq https://orcid.org/0000-0001-9275-6902
Mohammad Sarfraz Electrical Engineering Department, Aligarh Muslim University, Aligarh, India https://orcid.org/0000-0002-5586-3546
Muhammad Mazin Al Maathidi Computer Science Department, University of Bahrain, Bahrain https://orcid.org/0000-0002-3328-5474

DOI:

https://doi.org/10.52866/ijcsm.2024.05.02.008

Keywords:

lie speech, machine learning, voice stress analysis, Real time lie detection, Random Forests

Abstract

Lie detection has gained importance and is now extremely significant in a variety of fields. It plays
an important role in several domains, including law enforcement, criminal investigations, national security,
workplace ethics, and personal relationships. As advances in lie detection continue to develop, real-time
approaches such as voice stress technology have emerged as a feasible alternative to traditional methods such as
polygraph testing. Polygraph testing, a historical and generally established approach, may be enhanced or replaced
by these revolutionary real-time techniques. Traditional lie detection procedures, such as polygraph testing, have
been challenged for their lack of reliability and validity. Newer techniques, such as brain imaging and machine
learning, might offer better outcomes, although they are still in their early phases and require additional testing.
This project intends to explore a deception-detection module based on sophisticated speech-stress analysis
techniques that might be applied in a real-time deception system. The purpose is to study stress and other
articulation cues in voice patterns, to establish their precision and reliability in detecting deceit, by building upon
previous knowledge and applying state-of-the-art architecture. The performance and accuracy of the system and its
audio aspects will be thoroughly analyzed. The ultimate purpose is to contribute to the advancement of more
accurate and reliable lie-detection systems, by addressing the limitations of old techniques and proposing practical
solutions for varied applications. This paper proposes an efficient feature-selection strategy, which uses random
forest (RF) to select only the significant features for training when a real-life trial dataset consisting of audio files
is employed. Next, utilizing the RF as a classifier, an accuracy of 88% is reached through comprehensive
evaluation, thereby confirming its reliability and precision for lie-detection in real-time scenarios.