Cross-cultural production and detection of deception from speech

Detecting deception from different dimensions of human behavior has been a major goal of research in psychology and computational linguistics for some years and is currently of considerable interest to military and law enforcement agencies. However, relatively little work has been done to develop automatic methods to detect deception from spoken language or to compare deception detection and production between different cultures. We present results of experiments on a new corpus of deceptive and non-deceptive speech, collected from native speakers of Standard American English and Mandarin Chinese, all speaking English, to investigate acoustic, prosodic, and lexical cues to deception. We report first on the role of personality factors derived from the NEO-FFI (Neuroticism-Extraversion- Openness Five Factor Inventory) and of gender, ethnicity and confidence ratings on subjects' ability to deceive and to detect deception. We then present classification results discriminating deceptive from non-deceptive speech, using these features as well as acoustic and prosodic cues. We find that combining acoustic and prosodic features with information about the speaker's personality, gender, and language results in a classification accuracy of 65.86%, which represents ~10% relative improvement from baseline accuracy.