Real-Time Speaker Detection for User-Device Binding

摘要

This thesis explores the accuracy and utility of a framework for recognizing a speaker by his or her voice called the Modular Audio Recognition Framework (MARF). Accuracy was tested with respect to the MIT Mobile Speaker corpus along three axes: 1) number of training sets per speaker, 2) testing sample length and 3) environmental noise. Testing showed that the number of training samples per speaker had little impact on performance. It was also shown that MARF was successful using testing samples as short as 1000ms. Finally, testing discovered that MARF had difficulty with testing samples containing significant environmental noise. An application of MARF, namely a referentially-transparent calling service, is described. Use of this service is considered for both military and civilian applications, specifically for use by a Marine platoon or a disaster-response team. Limitations of the service and how it might benefit from advances in hardware are outlined.