Speech recognition by dynamic recurrent neural networks

In this paper, a dynamic recurrent neural network (RNN) is trained to recognize speech data and analyzed its ability of recognition. The results of this analysis indicate that the RNN has the capability to spot a specific word in any length of connected words, even if they are spoken by unknown speakers. Furthermore, this paper shows that the degenerated attractors and the transient states of the autonomous system, which is determined by each of the input vectors, is important to explain the facilities of the RNN.