| Detecting speech under stressful conditions can have broad ramifications in many applications that require rapid response to avert undesirable outcomes. The speech recognition system must be capable of detecting key words spoken under stress and provide a basis for any immediate corrective actions that can be taken. Examples of this include panic calls from cellular phone users and drivers on busy highways, cockpit commands of pilots in commercial/fighter planes, emergency calls for medical help, and so on. In this context, this dissertation presents a framework for the development of a Speech Command and Control System (SCCS) capable of detecting words spoken under stress.;Typical methods to detect stress in speech include measuring skin conductivity, detecting changes in body temperature, and measurement of changes in heart rate of the subject. These measurements require that the subject be physically connected to monitoring systems thereby placing constraints on the subject's mobility. A successful SCCS, on the other hand, would eliminate any physical connections and allow the subject to carry out other useful tasks that may be necessary during stressful and/or demanding situations.;In this research, the feasibility of a SCCS is demonstrated through a simulation model. The simulation model incorporates the recognition of specific commands that are needed during tests performed at the Aerial Cable Range (ACR) of the White Sands Missile Range Test Facility. A speech editor is developed to record speech commands, extract and analyze speech features, classify commands as either normal or under stress, and activate the desired functions of the ACR simulation model. Linear Predictive Coding (LPC) parameters are used to train a neural network-based speech classifier system. Results indicate that a successful SCCS can be developed using a neural network-based approach. |
