Speaker's independent and isolated word voice recognition system using neural network

PURPOSE: To obtain a method for recognizing a speaker's non-dependent and isolated word voice. CONSTITUTION: The method based on a mixed recognition system uses a neural network, its parallel processing can be used for improving recognition in time and memory and optimizing the system and some of the integrated side of recognition technology is also maintained. Each complete word is modeled by a left or right Markov model automaton having recurrence to a state corresponding to the acoustic part of the word. Recognition is obtained by applying dynamic algorithm based on Viterbi algorithm to all automatons for detecting a minimum cost route corresponding to a recognized word, an emission probability is calculated through a neural network having feedback trained by a unique method and transition probability is estimated by a suitable method. COPYRIGHT: (C)1994,JPO