Can Non-Linear Readout Nodes Enhance the Performance of Reservoir-Based Speech Recognizers?

It has been shown for some time that a Recurrent Neural Network (RNN) can perform an accurate acoustic-phonetic decoding of a continuous speech stream. However, the error back-propagation through time (EBPTT) training of such a network is often critical (bad local optimum) and very time consuming. These problems hamper the deployment of sufficiently large networks that would be able to outperform state-of-the-art Hidden Markov Models. To overcome this drawback of RNNs, we recently proposed to employ a large pool of recurrently connected non-linear nodes (a so-called reservoir) with fixed weights, and to map the reservoir outputs to meaningful phonemic classes by means of a layer of linear output nodes (called the readout nodes) whose weights form the solution of a set of linear equations. In this paper, we collect experimental evidence that the performance of a reservoir-based system can be enhanced by working with non-linear readout nodes. Although this calls for an iterative training, it boils down to a non-linear regression which seems to be less critical and time consuming than EBPTT.