Date of Award

Summer 1992

Document Type


Degree Name

Doctor of Philosophy (PhD)


Electrical & Computer Engineering


Electrical Engineering

Committee Director

Stephen A. Zahorian

Committee Member

David Livingston

Committee Member

C, Michael Overstreet

Committee Member

John Stoughton


This dissertation introduces a binary partitioned approach to statistical pattern classification which is applied to talker identification using neural networks. In recent years artificial neural networks have been shown to work exceptionally well for small but difficult pattern classification tasks. However, their application to large tasks (i.e., having more than ten to 20 categories) is limited by a dramatic increase in required training time. The time required to train a single network to perform N-way classification is nearly proportional to the exponential of N. In contrast, the binary partitioned approach requires training times on the order of N2. Besides partitioning, other related issues were investigated such as acoustic feature selection for speaker identification and neural network optimization.

The binary partitioned approach was used to develop an automatic speaker identification system for 120 male and 130 female speakers of a standard speech data base. The system performs with 100% accuracy in a text-independent mode when trained with about nine to 14 seconds of speech and tested with six to eight seconds of speech.