Noisy speech recognition based on modified RBF network

Speech recognition rate is often influenced by noise because of mismatching between training and recognition environments. This paper present a recognition system based on modified RBF Network. In this system, training data are predistorted by morphology filter same as recognizing data. For modified RBF network, cluster centers are decided by competitive learning algorithm, weights are achieved using conjugate gradient descent algorithm, network structure is optimized with pruning hidden neurons, the parameters of the hidden layer are trained dynamically. Experiments show that this system can increase system performance in noisy environment.