Speaker independent isolated word recognition system using neural networks

Abstract From the Canadian patent

A method for speaker independent isolated word recognition is based on a hybrid recognition system using neural networks and parallel processing to give improved recognition at optimized speed and memory use while also keeping some of the consolidated aspects of recognition techniques. Complete words are modeled with left-to-right Markov model automata with recursion on states each of which corresponds to an acoustic portion of the word. Recognition is obtained by performing dynamic programming according to the Viterbi algorithm on all automata in order to detect the one having the minimum cost path correspon- ding to the recognized word. Emission probabilities are computed through a neural …