Time-state neural networks (TSNN) for phoneme identification by considering temporal structure of phonemic features

The author reports a new structure for phoneme identification neural networks, time-state neural networks (TSNNs). Phonemes in Japanese have certain rough temporal structures which do not greatly change even when the utterance is an isolated word or continuous speech. Thus, if the neural network is to treat this, it would be very helpful to be able to identify phonemes whatever the utterance. The proposed TSNNs are able to deal with the temporal structure of phonemic features. Several types of TSNNs are described along with their phoneme identification performance, tested on the Japanese phonemes /b,d,g,m,n,N/ taken from isolated word, phrase, and sentence utterances. The phoneme identification performance of the TSNNs was much better than that of conventional TDNNs, especially in the case of continuous speech