Experiments on the implementation of recurrent neural networks for speech phone recognition

This paper reports on an extensive set of experiments that explore training methods and criteria for recurrent neural networks (RNNs) used for speech phone recognition. Seven different criterion functions are evaluated for speech recognition. A new criterion function that allows direct minimization of the frame error rate is proposed. Two new optimization methods for RNN weight updating are investigated. Experiments have been carried out on the Intel Paragon parallel processing system. The performance of the resulting phone recognition system is competitive with the best results in the literature.