Conductance-Threshold Dual Adaptive Spiking Neural Networks for Speech Recognition

Spiking neural networks (SNNs) have become popular in brain-like computing due to their biological plausibility and computational power. SNNs use spike coding to integrate temporal and frequency characteristics of information, making them advantageous for processing dynamic time-series signal-related problems such as speech signals. However, to effectively simulate the neural system and solve practical problems, it is crucial to construct a suitable spiking neuron model with realistic physiological properties and computational efficiency. In this work, we proposed a dual adaptive Integrate-and-Fire neuron model (DAIF) with dynamic conductance and threshold, which has more biologically realistic dynamics and is relatively simple to compute. Based on this model, a recurrent neural net-work for speech recognition was constructed using different spiking neuron models, leading to the establishment of a complete speech recognition model. We also conducted simulation experiments on the DAIF model and tested it on the Spike Heidelberg Digits (SHD) dataset, yielding good experimental results. Our work highlights the potential value of constructing efficient spiking neuron models for speech recognition tasks.