基于方位特征的听觉选择性注意计算模型研究

发布时间：2018-08-27 06:48

【摘要】：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.
[Abstract]:Classical auditory attention calculation models mainly focus on the primary auditory characteristics such as sound intensity, frequency, time and so on. These characteristics can not well simulate the directionality of auditory attention, so we must seek more advanced auditory features to distinguish different sounds. According to auditory perception mechanism, a bottom-up auditory selective attention computing model is proposed based on acoustic source azimuth characteristics and neural network. The model firstly preprocesses and analyzes the spectrum of binaural signals, and then sends them into the where path and the what path respectively, in which the where path is used to extract the azimuth characteristic parameters, and the local azimuth characteristics of the sound source are extracted by using the neural network. Then the azimuth feature saliency map is obtained by local feature aggregation and global optimization. Finally, the dominant azimuth is extracted from the azimuth feature saliency map and acted on the what path, and the corresponding dominant sound is separated by time-frequency masking method. The simulation results show that the azimuth feature is introduced into the model as the clustering clue, and the multi-level neural network is used to automatically screen the sound object that deserves attention, and the dominant sound in the complex acoustic environment is extracted in real time. The orientation classification mechanism, attention selection mechanism and attention transfer mechanism of human hearing are well simulated.
【作者单位】：四川大学电子信息学院;
【分类号】：TN912.3;TP183

【参考文献】