Bilkent University
Department of Computer Engineering
CS 590/690 SEMİNAR

 

Attentional Deceit Detection using Audio Spectrograms

 

Maria Raluca Eskin

Master Student
(Supervisor: Asst.Prof.Dr.Hamdi Dibeklioğlu )
Computer Engineering Department
Bilkent University

Abstract: Deception occurs commonly in our day-to-day life, yet people are inapt to capture this phenomenon in natural settings. Thus, automated detection of deceptive intent is an important task. This work addresses the complex problem of deceit detection by exploring various attention proposals for vocal modeling of speech. To this end, we utilize 2-dimensional spectrograms, extracted at a constant rate, to obtain specific human vocal patterns from each audio input. Our method employs a convolutional recurrent network enhanced with both spatial and temporal attention methods. The model takes the input spectrograms from successive timesteps and outputs a single latent representation encapsulating the deceptive content of the whole audio clip. In this talk, I will present the current status of our research and a contrastive view on the effect of various attention proposals on classification performance achieved with the popular high-stakes, in-the-wild, multimodal Real-Life Trial (RLT) dataset.

 

DATE: 12 December, Monday @ 15:30 Place: EA 502