An important organizing principle observed in the sensory pathways in the brain is the orderly placement of neurons. Although the neurons are structurally identical, the specialized role played by each unit is determined by its internal parameters that are made to change during early learning processes. In the human auditory system, the nerve cells and fibres are arranged in a manner that would elicit maximum response from the neurons when they are activated. Although most of this organization is genetically determined, some of the high level organization is created due to algorithms that promote self-organization. Kohonen's self-organizing feature map is a neural net model that produces feature maps similar to the ones produced in the brain. These maps are capable of describing topological relationships of input signals using a one or two dimensional representation. This technique uses unlabeled data and requires no training as in supervised learning algorithms. It is hence immensely useful in speech and vision applications. This neutral net has been implemented for the recognition of vowels in the American English language. The net has been trained and tested with vowel data. The formation of internal clusters or categories has been observed and closely reflects the tonotopic relationships between the vowels. An analysis of the results has been carried out and the performance has been compared to other classification techniques. A graphical user interface has also been developed using Xview to help visualize the formation of the maps during the training and testing processes.
Library of Congress Subject Headings
Speech processing systems--Design and construction; Automatic speech recognition; Pattern recognition systems--Design and construction; Vowels--Analysis--Data processing
Department, Program, or Center
Computer Science (GCCIS)
Sundaram, Anand R. K., "Vowel recognition using Kohonen's self-organizing feature maps" (1991). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus