华南理工大学学报(自然科学版) ›› 2007, Vol. 35 ›› Issue (9): 20-25.

• 电子、通信与自动控制 • 上一篇    下一篇

利用神经网络外推中垂面上低仰角HRTF 的方法

钟小丽   

  1. 华南理工大学 物理科学与技术学院,广州 广东 510640
  • 收稿日期:2006-10-12 出版日期:2007-09-25 发布日期:2007-09-25
  • 通信作者: 钟小丽(1975-) ,女,博士,讲师,主要从事音频信号处理和电声技术方面的研究. E-mail:xlzhong@ scut.edu. cn
  • 作者简介:钟小丽(1975-) ,女,博士,讲师,主要从事音频信号处理和电声技术方面的研究.
  • 基金资助:

    国家自然科学基金资助项目(1 0374031 )

Method to Extrapolate HRTF at Low Elevation of Median Plane by Means of Neural Network

Zhong Xiao-li   

  1. School of Physical Science and Technology , South China Univ. of Techn. , Guangzhou 510640 , Guangdong , China
  • Received:2006-10-12 Online:2007-09-25 Published:2007-09-25
  • Contact: 钟小丽(1975-) ,女,博士,讲师,主要从事音频信号处理和电声技术方面的研究. E-mail:xlzhong@ scut.edu. cn
  • About author:钟小丽(1975-) ,女,博士,讲师,主要从事音频信号处理和电声技术方面的研究.
  • Supported by:

    国家自然科学基金资助项目(1 0374031 )

摘要: 为了构建关于仰角方向连续的头相关传输函数(HRTF) ,提出采用径向基函数神经网络外推中垂面上低仰角方向HRTF 的方法.基于KEMAR 人工头和真人的数据,对3种不同网络输入方式下外推方法的有效性进行了研究.结果表明:外推值和测量值的相关系数可达0.93; 增大网络输入的数据量、减小外推方向和已知方向的空间距离可提高外推的准确性;外推的误差随频率的升高而增大.

关键词: 头相关传输函数, 外推, 神经网络, 虚拟现实

Abstract:

In order to construct a continuous head-related transfer function (HRTF) along elevation , a method to extrapolate the HRTF at low elevation of the median plane is proposed based on the radial basis function (RBF) neural network. Then , according to the data of KEMAR mannequin and a human being , the performances of the proposed method in three network input methods are analyzed. The results show that the correlative coefficient between he extrapolated and the measured HRTFs is as high as O. 93 , that the extrapolation accuracy can be improved by increasing the data of network input and by reducing the spatial distance between the extrapolated and the known directions , and that the extrapolation error increases with frequency.

Key words: head-related transfer function, extrapolation, neural network, virtual reality