A probabilistic training scheme for the time-concentration network

Abstract

The present paper studies the training aspect of the time-concentration neural network in the context of automatic speech recognition. The original model [Tank and Hopfield, 1987], uses constant connection weights to perform word-spotting from a input grapheme string. However, a more robust and automatic assignment of connection weights (i.e. training) is required for complex tasks such as speech recognition. A generalised training scheme based on probabilistic formulation has been proposed to enhance the performance of the network for a speech recognition experiment. Improvement of performance has been acheived using the proposed modification over the original formulation.

uunet!shakti!tifr!csc11