Identifying Deceptive Speech Across Cultures
Abstract
We have completed our collection of deceptive and non-deceptive speech recorded from interviews between native speaker of Mandarin and of English instructed to answer truthfully or to lie about 24 biographical questions. Subjects were rewarded or penalized financially for their ability to lie (interviewee) or to distinguish truth from lie (interviewer); each subject acted both roles. At 125h (174 subjects), this is by far the largest cleanly recorded speech corpus of its kind. From this data, we find that ability to lie is significantly correlated with ability to detect deception. We also find significant correlations of deception ability with personality factors (extraversion, conscientiousness). Using acoustic-prosodic features, gender, ethnicity and personality information our machine learning experiments can classify truth vs. lie in our data with 65 accuracy; we expect better results when we include lexical features.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 25, 2016
- Accession Number
- AD1012588
Entities
People
- Julia B. Hirschberg