Speech Based Deception Detection Using Bispectral Analysis

Abstract

Speech is considered as one of the most efficient and effective way to communicate with each other. However, a deception is a very common phenomenon in speech communication. It is difficult to detect if anyone is actually telling the truth or not. This study proposes a neural response based novel technique to identify the true or false from speech. In this study, the speech signal is used as the input to the auditory nerve model. This technique applies the higher order statistics called bispectrum to the auditory neurogram to distinguish the true and false from speech. Different parameters of the bispectrum are used as a feature to detect a deception from speech. Deceptive speech can be detected accurately by using the 'normalized bispectral entropy' of the bispectrum feature parameters for the envelope information (ENV) data and the 'maximum bispectrum' of the bispectrum feature parameter for the temporal fine structure (TFS) data. Speech based deception detection is a speech processing method which provides better accuracy to detect deception than many other deception detection techniques. This technique could be applied effectively for the national security systems.