Yasuhiro KOMORI, Shigeki SAGAYAMA, Alexander H. WAIBEL
A Fuzzy Training Approach
for Phoneme Classification
Neural Networks
Abstract:In this report, we propose a new fuzzy training approach for a phoneme classification type neural networks. This fuzzy training approach is realized through back-propagation algorithm, but differs from the conventional training approach in the point of how to give the training target values for the neural networks. In the conventional training approach, the phoneme class of the input data are given to the target values for training; 1 for the output unit which corresponds to the input phoneme, and 0 for the other output units. However, in this fuzzy training approach, the target values are defined as how likely the input phoneme is to the phoneme classes. This likelihood is computed according to the distance between the input phoneme itself and other data in training data set. The phoneme classification experiments are performed on Japanese/bdgmnN/, 14 category English vowels and 40 category English all phonemes. This report also discussed these experiments.