Scholarlink

<p>The Head-Related Transfer Function (HRTF) is essential for creating immersive audio environments. However, HRTF measurement systems typically capture HRTFs sparsely for an individual, and sparse HRTFs cannot meet the demands of virtual acoustic applications. Therefore, spatial interpolation of HRTFs is necessary to obtain impulse responses (IRs) for various orientations. This paper proposes a method using a Physics-Informed Autoencoder (PIAE) to perform spatial interpolation on sparse HRTFs. The PIAE utilizes the Helmholtz equation (the governing equation of sound wave propagation) to regulate the spatial interpolation process, which helps in generating physically valid data. However, using the Helmholtz equation alone cannot generate satisfactory data, especially for unknown locations. Therefore, it is necessary to use source-location-independent representations to guide the model in generating realistic data. The numerical experiments show that, compared to methods such as source-position-conditioned autoencoders, Generative Adversarial Networks (GAN), transfer learning, and Physics-Informed Neural Networks (PINN), the method proposed in this paper shows superior performance when the spatial orientation is sparse.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Spatial interpolation of head-related transfer functions using a physics-informed autoencoder

  • 期刊论文

  • 发表日期：

    2025年5月12日

查看全文
  • Wei Chen,
  • Xiaogang Wei

摘要

The Head-Related Transfer Function (HRTF) is essential for creating immersive audio environments. However, HRTF measurement systems typically capture HRTFs sparsely for an individual, and sparse HRTFs cannot meet the demands of virtual acoustic applications. Therefore, spatial interpolation of HRTFs is necessary to obtain impulse responses (IRs) for various orientations. This paper proposes a method using a Physics-Informed Autoencoder (PIAE) to perform spatial interpolation on sparse HRTFs. The PIAE utilizes the Helmholtz equation (the governing equation of sound wave propagation) to regulate the spatial interpolation process, which helps in generating physically valid data. However, using the Helmholtz equation alone cannot generate satisfactory data, especially for unknown locations. Therefore, it is necessary to use source-location-independent representations to guide the model in generating realistic data. The numerical experiments show that, compared to methods such as source-position-conditioned autoencoders, Generative Adversarial Networks (GAN), transfer learning, and Physics-Informed Neural Networks (PINN), the method proposed in this paper shows superior performance when the spatial orientation is sparse.