Stress Recognition Using Sparse Representation of Speech Signal for Deception Detection Applications in Indian Context

This work aims at stress recognition of a person using sparse representation of speech samples for a real time application of deception detection. A learned dictionary using K-Singular Value Decomposition (K-SVD) is considered for sparse signal representation. The work incorporates parameterization of nonlinear Teagor Energy Operated Critical Band Autocorrelation Envelop (TEO-CB-Auto-Env) feature extraction of speech samples. The database developed by IIT Guwahati with five stress categories includes Angry, Happy, Lombard, Neutral and Sad. The corpus has speech samples in Hindi language and hence the proposed stress classification system is applicable for Indian context. Weka, Neural Network and Autoencoder Neural Network are used for classification. A better recognition accuracy around 5% higher is gained in this work compared with earlier work on speech based stress recognition on the same database. The comparative performance shows that Random Forest classifier from Weka performs better than Neural Network and Autoencoder Neural Network.