Speech Recognition Model with Spatio-Temporal Encoding and Brain-Like Neural Networks

Deep learning has achieved great success in various applications, such as image processing and audio signal processing. However, deep learning methods usually require large and powerful computing resource, which greatly limit the practical application of deep learning methods. The spiking neural network is a new paradigm of neural networks and has the potential for low-power consumption. In this paper, we proposed a new computational model with the aim to provide an efficient and low-power computational model for speech recognition tasks. In the proposed model, the threshold coding method is employed to transform the speech signals to spike train patterns. The encoded spike patterns are then presented to the designed multi-layer spiking neural networks and trained by the surrogate-learning algorithm. After learning, the classification is decided by the firing rate of the neurons in the last layer. Experimental results demonstrate that the proposed method gets good accuracy in the speech recognition task. In addition, the proposed computational model works in an event-driven manner, hence, it holds great potential for low-latency and low-power consumption in neuromorphic hardware.