Detection of time-pressure induced stress in speech via acoustic indicators

Matthew Frampton,  Sandeep Sripada,  Ricardo Augusto Hoffmann Bion,  Stanley Peters
Stanford University


Abstract

We use automatically extracted acoustic features to detect speech which is generated under stress, achieving 76.24% accuracy with a binary logistic regression. Our data are task-oriented human-human dialogues in which a time-limit is unexpectedly introduced partway through. Analysis suggests that we can detect approximately when this event occurs. We also consider the importance of normalizing the acoustic features by speaker, and detecting stress in new speakers.