A comparison between recurrent neural architectures for real-time nonlinear prediction of speech signals

This paper presents a comparative study on the performance of recurrent neural networks trained in real-time to predict the next sample in a speech signal. The comparison is basically done versus linear predictors, and a pipelined recurrent neural network which has been proposed for this task. Results confirm those of previous works where limitations to deal with numeric time series were detected for recurrent neural architectures, specially when using the real-time recurrent learning algorithm. The decoupled extended Kalman filter training algorithm, on the other hand, overcomes partially some of these limitations.