基于稀疏表示和神经网络的头相关传输函数个性化方法研究

Research on personalized modeling of head-related transfer function based on sparse representation and neural network

头相关传输函数(Head-Related Transfer Function, HRTF)的个性化定制,是实现虚拟听觉系统(Virtual Audio Display, VAD)的关键技术之一。本文提出了一个基于稀疏表示和径向基函数(adial Basis Function, RBF)神经网络的HRTF个性化方法,通过LASSO回归分别计算出生理特征的稀疏系数和HRTF数据的稀疏系数,利用神经网络来建模两组系数的映射关系,并使用Pearson相关分析筛选与测试样本相关性强的数据作为训练集,所提方法只需要进行较少的训练就可以估计出个性化头相关传输函数。仿真实验表明,与已有的稀疏表示方法相比,本方法所需的训练集更小,估计误差更低。

The synthesis of personalized head-related transfer function (HRTF) is a key factor in virtual audio display. In this paper, a method based on sparse representation and radial basis function (RBF) neural network is proposed to obtain personalized HRTF, to construct a personalized model, LASSO regression is applied to get the sparse representations of subject’s anthropometric features and subject’s HRTF data respectively;then a neural network is used to model the relationship between the two representation sets. And some important samples are selected as training set according to the Pearson correlation analysis. Experiments show that our method outperforms the previous sparse representation method with smaller train set and lower estimation error.