MIT Media Lab, E14-633
Human nonverbal behaviors are subtle, fleeting, and often contradictory. Is it possible for computers to not only sense, translate and interpret human nonverbal behaviors, but also help us improve these behaviors? This thesis presents a computational framework and a user-centric evaluation to answer that question.
The core of this thesis contains three main technical components:
Automatically capture the nonverbal nuances from face and speech using computer vision and machine learning techniques.
Respond naturally in real time to human behaviors using an autonomous embodied character.
Interpret and represent conversational multimodal behavioral data into an intuitive and educational format using data visualization techniques.
To validate the research hypothesis, MACH (My Automated Conversation coacH), an embodied 3D character, was designed. MACH is able to “see”, “hear” and “respond” in real-time through a webcam and a microphone using an ordinary laptop. The experiment was contextualized for job interviews where MACH played the role of the interviewer, asked interview questions, and at the end, provided feedback. The effectiveness of MACH was assessed through a weeklong trial with 90 MIT undergraduates. Students who interacted with MACH were rated by human experts to have improved in overall interview performance, expressing excitement about the job, and were more likely to be recommended for the position, while the ratings of students in control groups did not improve.
Findings from this thesis could open up new interaction possibilities of helping people with public speaking, social-communicative difficulties, language learning or even dating!
Host/Chair: Rosalind W. Picard
Jeffrey Cohn, Bilge Mutlu, Louis-Philippe Morency