Phoneme recognition with statistical modeling of the prediction error of neural networks

This paper presents a speech recognition system which incorporates predictive neural networks. The neural networks are used to predict observation vectors of speech. The prediction error vectors are modeled on the state level by Gaussian densities, which provide the local similarity measure for the Viterbi algorithm during recognition. The system is evaluated on a continuous speech phoneme recognition task. Compared with a HMM reference system, the proposed system obtained better results in the speech recognition experiments.