Speech based automatic lie detection

Abstract

This work studies the effect of the emotions that is experienced due to a guilt situation on different vocal parameters in an attempt to identify whether or not the suspect is lying. The homomorphic speech processing is applied to extract the vocal parameters related to the source excitation such as: pitch, pitch power and vowel duration and those related to the vocal tract such as: formant frequencies and its gain. Also the energy, as a global vocal parameter, is computed. The vocal parameters are extracted from normal speech utterances and from stressed utterances for the same suspect in order to determine the most significant vocal parameters that can be affected by emotional stress. The correlation coefficients were investigated between the pitch power and the digitized smoothed output of the psychological stress evaluator (PSE). More than a 0.8 correlation coefficient has been found. Six cases of real time criminal suspects cases were investigated throughout this work. Traditional lie detection questioning techniques were used to develop questionnaires for these criminal cases. Moreover, a case for an actor simulating different emotional states (downloaded from the Internet) was investigated for the effect of different emotions on the vocal parameters. Speech vocal parameters and the PSE Hirch & Wiegele (1981) scoring method were investigated for stress (due to anxiety or guilt). The pitch contour exhibits the most significant sensitivity for speech-based stressed/unstressed classification.