Linear and nonlinear speech feature analysis for stress classification

There are many stressful environments which deteriorate the performance of speech recognition systems. Examples include aircraft cockpits, 911 emergency telephone response, high workload task stress, or emotional situations. To address this, we investigate a number of linear and nonlinear features and processing methods for stressed speech classi cation. The linear features include properties of pitch, duration, intensity, glottal source, and the vocal tract spectrum. Nonlinear processing is based on our newly proposed Teager Energy Operator (TEO) speech feature which incorporates frequency domain critical band lters and properties of the resulting TEO autocorrelation envelope. In this study, we employ a Bayesian hypothesis testing approach and a hidden Markov model (HMM) processor as classi cation methods. Evaluations focused on speech under loud, angry, and the Lombard e ect1 from the SUSAS database. Results using receiver operating characteristic (ROC) curves and EER (equal error rate) based detection show that pitch is the best of the ve linear features for stress classi cation; while the new nonlinear TEO-based feature outperforms the best linear feature by +5.2%, with a reduction in classi cation rate variability from 8.66 to 3.90.