Analysis of emotional stress in voice for deception detection

Abstract

The current work gives a method to estimate alterations in speech caused by stress. This estimation is used as an indicator to detect deceptive speech. Truthful or neutral speech and deceptive speech were elicited by forcing to answer "hot questions" related with politics and society. Data were processed by using log-likelihood ratios and Fisher's linear discriminant analysis. Independent results for male and female are presented. The classification results are around 100% for neutral speech in both genders, while the best classification rate (67%) for stressed speech is achieved for females. In a first approach we have seen, that subjects tend to be more deceptive when the questions are related with gender issues, giving a politically correct answer instead of their true opinion.