Extreme Learning Layer: A Boost for Spoken Digit Recognition with Spiking Neural Networks

Abstract

The automatic recognition of speaker-independent digits requires high accuracy and robustness to noise and variability. Spiking neural networks (SNNs) are a promising model for this task, as they can mimic the temporal dynamics and energy efficiency of the human auditory system. However, SNNs are difficult to train and often require complex learning algorithms. Spoken digits provide a useful benchmark task to evaluate new SNN architectures. Performance in small vocabulary tasks is an important first step before scaling up to more complex recognition scenarios. In this paper, we propose to use an extreme learning layer (ELL) as a simple and effective way to improve the learning of SNNs for spoken digit recognition. The ELL is a randomly generated layer that maps the input features to the next layer without any further adjustment. The output layer is then trained by entropy minimization. We show that ELL can boost the performance of the SNN on the benchmark data set TIDIGITS. We also compare our approach with some state-of-the-art methods achieving competitive results with less computational cost and complexity. The proposed approach also shows good robustness to additive noise.