A comparison between spiking and differentiable recurrent neural networks on spoken digit recognition

In this paper we demonstrate that Long Short-Term Memory (LSTM) is a differentiable recurrent neural net (RNN) capable of robustly categorizing timewarped speech data. We measure its performance on a spoken digit identification task, where the data was spike-encoded in such a way that classifying the utterances became a difficult challenge in non-linear timewarping. We find that LSTM gives greatly superior results to an SNN found in the literature, and conclude that the architecture has a place in domains that require the learning of large timewarped datasets, such as automatic speech recognition.