Deceptive Speech Detection Based on Sparse Representation

Generally, the extracted features of distinguishing deceptive speeches always focused on prosodic, vocal tract, lexical and glottal waveform features. The purpose of this paper is to examine the effectiveness of sparse coefficients for deception detection. In this paper, we firstly extract the Mel-Frequency Cepstrum Coefficient (MFCC) and Zero Crossing Rate (ZCR) from speech utterances as the input data of K-SVD algorithm to learn a mixture dictionary. And sparse coefficients are obtained by Orthogonal Matching Pursuit (OMP) algorithm. Then we use those coefficients as features to train Support Vector Machine (SVM) model and test the classifier accuracy based on the trained model. Finally, we present the experimental results of this approach and compare the results with the conventional features consisting of Short-Time, Pitch, Formant, and Duration based on corpus of Soochow University Speech Processing Researches-Deception Speech Detection Corpus (SUSP-DSD). It shows that sparse coefficients perform better than the conventional features in deception detection.