Speaker Classification based on Multiple Criteria

Department of Mechanical Engineering
Advisor: Professor V. Sundararajan

Abstract: This thesis presents techniques to classify speech signals produced by unfamiliar speakers into categories  based  on  criteria  such as  gender  and  age.    Such  classification  is  useful  to  augment  the efficiency of conventional video surveillance systems. Existing systems classify speakers along only one dimension.  This  thesis  develops  methods  to  classify  speakers  based  on  multiple  categories.  Age  group (children and adults) and gender group (male and female) are used  as  examples  to clarify these issues. Experimental evaluation indicates that 1) Speaker groups are dependent 2) The use of single classifier to sort  speakers  into  multi-criteria  groups  yields sub-optimal  results.  Thus,  a  sequential  approach  is necessary 3) The sequence of classification affects the performance of the system.