基于小波神经网络的HRTFs逼近与声频仿真

HRTFs APPROXIMATIONS AND AUDIO SIMULATION BASED ON WAVELET NEURAL NETWORKS

头部相关传递函数（HRTFs）是指从声源到耳鼓的谱滤波器，它因提取了声源的方位信息，所以在声频仿真中，它是非常重要的立体听觉定位曲线．由于HRTFs随声源的相对位置、频率和听觉对象不同而变化，并与其自变量之间还存在着非常复杂的非线性关系，所以王维声音仿真的实现需要处理庞大的HRTFs数据．提出了基于小波神经网络的HRTFs近似模型，即通过训练网络重新设置模型数据来获取空间任意位置的非个人HRTFs数据．声频仿真实验结果表明此逼近方法不仅保持了原始HRTFs的相关特征，而且还在三维声频仿真中提高了计算效率．

Head-related transfer functions (HRTFs) refer to the spectral filtering from sound sources to listeners' eardrums or ear canals. They are a very important cue to spatial hearing localization in 3-D audio simulation due to their ability of extracting sound source location information. Since HRTFs vary as a function of relative sound source locations, frequency, and hearing subjects, and at the same time, there exists a very complicated nonlinear relationship with its independent variables including azimuths, elevations, and frequency, practical implementation of 3-D audio simulation always faces a large set of HRTFs data. A kind of effective HRTFs nonlinear approximation model is presented based on wavelet neural networks to decrease a large set of data. The original HRTFs data can be reset by network training, thus non-individual magnitude HRTFs data at any position are obtained. Audio simulation experiment demonstrates that this HRTFs approximation model not only maintains perceptually relevant features of original HRTFs and but also improves computational efficiency for real-time implementation.