Speech recognition system with neural network

Claim 1

A voice recognition apparatus capable of recognizing any word utterance by using a neural network, said apparatus comprising: means for inputting an input utterance and for outputting compressed feature variables of said input utterance, said input means including means for receiving said input utterance, means connected to said receiving means for amplifying said input utterance, means connected to said amplifying means for extracting said feature variables from an electrical signal, and means connected to said extracting means for compressing said feature variables; a word-head detecting section for detecting a front endpoint of said input utterance from said compressed feature variables, said word-head detecting section outputting said compressed feature variables if said front endpoint is detected as a start end of said input utterance, said word-head detecting section not outputting said compressed feature variables if said front endpoint is not detected; a first means connected to said input means for receiving said compressed feature variables output from said word-head detecting section and for outputting a value corresponding to a similarity in partial phoneme series of a specific word among vocabularies to be recognized with respect to said input utterance, said first means being capable of sound analyzing said input utterance so that feature values are generated and shifted in a time scale and an input frame is selected so as to maximize each of said output values output from said first means corresponding to a similarity among said shifted feature values, said first means including a plurality of event nets for receiving feature variables extracted from an input utterance, each of said event nets being arranged to shift said feature variables, within a predetermined range from a front endpoint positioned at any time, in accordance with time interval information obtained by analyzing speech samples of a plurality of persons and by selecting a location at which a maximum output is made possible among shifted locations of said feature variables so that a value, corresponding to a similarity between said partial phoneme series of said corresponding word to be recognized and said input utterance, is output; a second means, connected to said first means, for receiving all of said values output from said first means and for outputting a value corresponding to a similarity in said specific word with respect to said input utterance; a third means, connected to said second means, for receiving all of said values output from said second means and for outputting a value corresponding to a classification of voice recognition, in which said input utterance belongs, so as to output a value corresponding to a similarity between said input utterance and words to be recognized.