In this paper, we present our new continuous speech recognition system, which is based on linked predictive neural networks, and a noise reduction neural network, which processes noise-collapsed speech signals in an auditory model based spectral domain. We also describe a series of noise reduction and speech recognition experiments in noise conditions in order to evaluate performance of both the LPNN-based recognizer and the noise reduction network. The experimental results show that our LPNN-based speech recognition system has better performance in noise conditions than a probabilistic model based system. In all experimental cases, the noise reduction network performed well and decreased the recognition error rate by at least 13%.

Keywords: Speech Recognition, Neural Networks, Noise Reduction.