The use of eigen-coded neural network for robust speech recognition

摘要

A multiclass classifier is used in any speech recognition system. The variations in the input speech along with the background noise Impose limitations on any practical speech classification system. Recently, some authors suggested the use of output-coded neural Networks to mitigate this problem. The basic idea is to use Codewords from an error-correcting code to label the different Classes. The neural network output is then decoded in the hope Of reducing classification errors. In the literature, the optimal Choice of an error-correcting code is not obvious. Some researchers Alluded to heuristic methods to choose a "good" code. In this Article, we derive an optimality condition, called "eigen" coding, for The choice of the code.