Isolated Word Recognition Using Fuzzy Set Theory

摘要

A unique algorithm was developed to recognize isolated words given the output of an extremely variable feature extraction process. Because of the high error rate of the acoustic processor, it was necessary to rely on the consistency of the sequences of phonemes, and the errors that typically occur for a given word to determine the word spoken. This was accomplished by generating error statistics and phoneme representations for each word in the vocabulary using a set of training speech files. The top five phoneme choices provided by the acoustic processor, and information indicating the accuracy of each choice, for each time segment of speech was implemented. Fuzzy set theory was used to combine this information with the error information obtained from the training files to determine the word spoken.