Recognition of Japanese phonemes using a neural network with template units.
Takakazu Shiomi, Kiyoshi Hashimoto · Journal of the Acoustical Society of Japan (E) · 1993
The multi-layer perceptron (MLP) trained by a back-propagation (BP) algorithm pro ducesa posterior probability. However, the MLP cannot reject input data of thecategories that the training data set does not include. On the other hand, if the threshold is defined, template-matching can reject the input data when the distance between input data and each template exceeds the threshold. In this paper, we present a template matchingneural network (TMNN), which has the characteristic of template-matching, and a Sub-Branch (SB) training algorithm. In order to confirm the effectiveness of the TMNN on phoneme recognition, we compared the result with the MLP, training by the BP algorithm.