Pattern recognition system and method using neural network

Claim 1

A speech pattern recognition system comprising: speech input means for inputting a speech pattern to be recognized; feature extracting means for extracting feature vectors from the input speech pattern; dictionary mean for storing a plurality of orthogonalized reference patterns, which are equal to eigenvectors calculated using Karhunen-Loeve (KL) expansion, and whose categories are known; inner product computing means, provided for each category, for computing inner products of the input speech pattern, using said feature vectors, whose category is unknown and the orthogonalized reference patterns of a category which are stored in said dictionary means and whose category is known; converting means, provided for each category, for nonlinearly converting the inner product of a category, which is computed by said inner product computing means, in accordance with a positive-negative symmetrical nonlinear function; evaluation value computing means, provided for each category and each evaluation value computing means including a multi-layer neural network whose elements each include multiplier means for multiplying an input signal from a lower adjacent layer by a preset weight coefficient and being connected to all elements constituting an upper adjacent layer, for computing an evaluation value of a category using the nonlinearly converted values calculated by said converting means and weight coefficients which are preset in said neural network of a category based on a discriminatively trained Back Propagation (BP) algorithm; comparing means for comparing the evaluation values obtained by said evaluation value computing means using said neural networks with other evaluation values; and discriminating means for outputting a category to which the input speech pattern belongs as a recognition result using comparison results of said comparing means.