Development of Spectral Speech Features for Deception Detection Using Neural Networks

Analysis of stress in speech signals offers a noninvasive alternative to detect deception. The aim of this study is to investigate and synthesize by exploring the spectral energy features namely delta energy and difference energy for deception detection from speech signals. The results are compared with innovative delta and time-difference cepstrum features. Extractions of the proposed speech features are based on the perception of human speech and the psychoacoustic masking property. To classify deceptive and truthful speech, spectral energy in twenty - one (21) bands of frequencies on a Bark scale is used. The reference speech signals for the database are garnered by using the track record of a blameworthy person throughout a police investigation. Two neural networks such as (1) the Levenberg-Marquardt (LM) algorithm and (2) the Long Short-Term Memory (LSTM) algorithm are used in the MATLAB environment to observe the performance and test the results. Simulation results indicate that the delta and time-difference energy are critical features to classify deceptive speech from truthful speech.