An integrated system for robust gender classification with convolutional restricted Boltzmann machine and spiking neural network

Different from traditional artificial neural networks (ANNs), spiking neural networks (SNNs) represent and transfer information in spikes, which are considered more like human brain. SNNs contain time information, which make them more suitable for addressing time-structured speech signals. However, it still remains challenging for spiking neural network (SNN) to implement classification tasks based on speech due to the lack of a proper encoding. In this paper, an integrated spiking neural network is proposed to perform the gender classification task. The whole system consists of three functional parts, including encoding, learning and readout. As convolutional restricted Boltzmann machine (CRBM) has shown outstanding capability for unsupervised learning of auditory features, we adopt it in this paper as a feature extractor, followed by a spike-latency encoding layer that converts the feature values into spike times. Then these spikes are processed by the spiking neural networks with the tempotron learning rule. We use the TIMIT database to evaluate the performance of our system. Our results show that the as-proposed system is robust for gender classification across a wide range of noise levels.