Robust Sound Recognition Based On Brain-Inspired Computing Method

In recent years,robust sound event recognition has attracted more and more researchers’ attention.In the real world,noise is everywhere.How to effectively and accurately identify the sound event in real-world noisy environment is still a challenging problem.However,feature extraction and model selection are the two most significant problems need to be solved in sound event recognition.The purpose of this study is to find an effective and feasible method of feature extraction and model selection to further improve the robustness of sound event recognition.Brain-inspired computing is a information processing method close to human brain.In this study,we explore the application of brain-inspired computing method to process the sound event recognition task and improve the robustness.Biological evidence suggested that the auditory systems utilize the local time-frequency regions with high signal-to-noise ratio(SNR)to process noise corrupted signals.Therefore,we propose a biologically realistic key-point detection technique to extract the local time-frequency information for robust sound event recognition by combining convolutional neural network,one of the most popularly applied methods in acoustic processing.Compared to traditional artificial neural network,spiking neural network contains time information which makes it more suitable for processing temporal-structured signals.Based on the advantage of processing temporal information,we proposed a system based on spike encoding and learning,which is used to identify and classify sound signals.The whole system contains encoding,learning and readout part.In the encoding part,we use key-point detection technique to extract sparse local time-frequency information.In a bid to improve the encoding performance,a dimension expansion is applied.Then a mapping operation is applied to obtain the spatio-temporal spike patterns.In the end,combining spiking neural network with tempotron learning rule for robust sound recognition.This paper uses the RWCP sound database for the experiments.The experimental results demonstrate that the as-proposed KP-CNN system can improve the performance of sound event recognition.The system based on spiking neural network can further improve the robustness of sound event recognition.