Construction and analysis of Indonesian-interviews deception corpus

Abstract

In this paper, we present the first deception corpus in Indonesian to support deception detection based on statistical machine learning approach due to the importance of data in related studies. We collect speech recordings along with their high frame rate video from 30 subjects to develop Indonesian Deception Corpus (IDC). Using financial motivation as its basic scenario, IDC consists of 5542 speech segments with a total duration of approximately 16 hours and 34 minutes. As an imbalanced corpus, the majority class is represented by truth segments which is almost four times higher than the lie segments. We also perform some experiments using only the speech corpus, along with the transcriptions. Using the combination of paralinguistic, prosodic, and lexical features, we obtained the best accuracy of 61.26% and F-measure of 61.30% using Random Forest classifier and RUS as the undersampling technique.