基于方位特征的听觉选择性注意计算模型研究

摘要

Classic computational model of auditory selective attention mainly involves simple characters such as intensity,frequency,and time,which cannot simulate directional auditory attention preferably and needs more advanced auditory features to the distinguish different source signals.According to the perception mechanism of auditory system,the paper presents a bottom-up auditory selective attention computational model that has two signal processing pathways involving orientation feature and neural network.In this model,firstly,the binaural signals are preprocessed,spectral analyzed and separately sent to dual pathways.The where pathway is used to extract parameters of orientation feature and local orientation features of signals with neural networks,Then the features are aggregated and globally optimized to gain a saliency map of orientation feature.Finally,the leading orientations are gained based on the saliency map and applied to the what pathway to separate leading signals by time-frequency masking.Orientation features are introduced as group clues in this module and multi-neural networks are used to extract objective signals from mix-signals automatically.Simulation results prove that the method proposed in this paper can dynamically extract leading signals from complex acoustic environments in real time,and that orientation classification,attention selection and attention switch of the human auditory system are well simulated.%经典的听觉注意计算模型主要针对声音强度、频率、时间等初级听觉特征进行研究,这些特征不能较好地模拟听觉注意指向性,必须寻求更高级的听觉特征来区分不同声音.根据听觉感知机制,本文基于声源方位特征和神经网络提出了一种双通路信息处理的自下而上听觉选择性注意计算模型.模型首先对双耳信号进行预处理和频谱分析;然后,将其分别送入where通路和what通路,其中where通路用于提取方位特征参数,并利用神经网络提取声源的局部方位特征,接着通过局部特征聚合和全局优化法得到方位特征显著图;最后,根据方位特征显著图提取主导方位并作用于what通路,采用时频掩蔽法分离出相应的主导音.仿真结果表明:该模型引入方位特征作为聚类线索,利用多级神经网络自动筛选出值得注意的声音对象,实时提取复杂声学环境中的主导音,较好地模拟了人类听觉的方位分类机制、注意选择机制和注意转移机制.