A New Method to Reduce the Ambiguity of Japanese Phoneme Candidates Recognized by Recurrent Neural Networks

Up to now, the method to reduce the ambiguity of phoneme recognition using 2nd-order Markov chain model of phonemes, has been proposed and has been evaluated by phonem lattice simulated and limited to substitution error. However, the method will be necessary to demonstrate the effectiveness for the phoneme candidate lattice obtained by actual speech recognition devices. This paper deals with recurrent neural networks(RNN) which are well- suited for natural language processing of speech recognition, specially for phoneme recognition. The ability of these networks has been investigated by phoneme recognition experiments using a number of Japanese words uttered by a native male speaker in a quiet environment. A method to detect the locations of devoicing vowels using the short- time average energy has been also proposed, and evaluated. Form results of the experiments, it is shown that recognition rates achieved with RNN are higher than those obtained with conventional non-recurrent neural networks, and that the method to detect the locations of devoicing vowels is useful.