Comparative Analysis of Classification Methods for Automatic Deception Detection in Speech

摘要

This paper presents the experimental results carried on the speech processing methods for paralinguistic analysis of deceptive and truthful statements. It includes a short survey of databases that contain both deceptive and truthful speech samples, as well as recently developed deception detection systems that were proposed within the framework of computational paralinguistic challenge ComParE-2016 and other scopes. Based on the analysis and comparison of different approaches for processing deceptive and truthful utterances the best methods and optimal parameters are reported as following. The highest performance in terms of Unweighted Average Recall (UAR) measure has been obtained by a Random Forest based classifier with UAR = 79.3%. High results have been shown by a single k-Nearest Neighbor classifier, as well as its combination with other classification methods such as Bagging and Classification via Regression, which demonstrated UAR = 76.3%.