Performance optimization of the echo state network for time series prediction and spoken digit recognition

In this article, the effect of the parameters for echo state network on the performance for time series prediction and spoken digit recognition is investigated. The results show that the normalized mean square error of the 30th-order NARMA time series is lower than 0.2 and the word error rate of the spoken digit recognition is lower than 0.03 when the reservoir size, reservoir sparsity, input weight scaling, and spectral radius of the reservoir connection matrix are selected reasonably.