A Modified Hilbert Analysis Method to Improve Voice Stress Analysis Systems

Abstract

Analyzing the cognitive load generated in the brain is the most important issue for specific applications such as voice stress analysis (VSA) systems in which the detection of stressed speech caused by an act of deception under law enforcement interview questioning or military interrogation. The most widely used algorithm for VSA systems is the empirical mode decomposition (EMD). Currently EMD that uses the cubic spline interpolation technique to find the envelopes of the non-periodic signal takes a long processing time, and to achieve accurate results the process is very time consuming and expensive otherwise some tests tend to produce inaccurate results. On the other hand EMD that uses Hilbert analysis method to speed up the process and provide more accurate results, suffer from finding the envelopes of the non-periodic signal. In this paper, a new algorithm is proposed for VSA, named fast Fourier transform (FFT) with a modified Hilbert analysis method (MH) for EMD algorithm, (FTT_MH_EMD), which provides a new technique that modifies the conventional Hilbert analysis method and combines it with the fast Fourier transform algorithm to overcome the previous limitations of using individually the FTT algorithm, the cubic spline interpolation technique or the conventional Hilbert analysis method and that can speed up the processing time and gives accurate results. Simulations and results witness that the proposed algorithm provides higher accuracy than the other attempts and also the processing time has dropped by 10 times faster than those in the products currently available in the market for VSA The advantages of using FFT for VSA are low cost when implemented in products and fast iterations method, and the disadvantage is the low resolution and therefore low accuracy (60%–70%) [1]. On the other hand the advantage of using Mcquiston Ford for VSA is the high accuracy results which are greater than 80%, and the disadvantage are complex algorithms and expensive implementation [5]. An important, barely audible, frequency is produced by the human vocal cords which could measure in a fairly accurate manner if the subject is being deceitful due to the effect the brain’s cognitive load has on the vocal cords [1 and 5]. This frequency is called the micro-tremor frequency generated by the involuntary micro-tremor muscle fibers in the vocal cords. When the subject is being deceitful, more work is induced on the brain which as a result causes tremor muscles to tense up in a non-voluntary manner [1 and 6]. Different non voluntary muscles would tense up including the micro-tremor muscle fibers found on the vocal cords. This induces a higher micro-tremor frequency range notably showing a large quantity of induced cognitive load and therefore a deceitful subject [7]. The micro-tremor frequency of all muscles in the body, including the vocal chords, vibrates in the 8 to 12 Hz range in the normal case without any stress, and that the VSA vendors claim to be the sole source of detecting if an individual is lying or is under stress [1, 4 and 8]. The paper is organized from seven sections EMD, EMD with Hilbert Analysis, previous work and limitations, the proposed algorithm, the proposed VSA system model, results and simulations and finally the conclusion