Acoustic-phonetic decoding based on Elman predictive neural networks

We present a phoneme recognition system based on Elman predictive neural networks. The recurrent neural networks are used to predict the observation vectors of speech frames. Recognition of phonemes is done using the prediction error as a distortion measure in the Viterbi algorithm. The performance of the neural predictive networks is evaluated on both the training database and on a speaker independent test database. The results obtained on the training database are similar to a four state continuous density HMM, results on the test database results are comparable to a three state HMM