Abstract
Many methods have been proposed for modeling head-related transfer functions (HRTFs) and yield a good performance level in terms of log-spectral distortion (LSD). However, most of them utilize linear weighting to reconstruct or interpolate HRTFs, but do not consider the inherent nonlinearity relationship between the basis function and HRTFs. Motivated by this, a domain knowledge-assisted nonlinear modeling method is proposed based on bottleneck features. Domain knowledge is used in two aspects: to generate input features derived from the solution to sound wave propagation equations at the physical level, and to design the loss function for model training based on objective evaluation criteria (LSD). Furthermore, by utilizing the strong representation ability of bottleneck features, the nonlinear model achieves a more accurate mapping. Objective and subjective experimental results show that the proposed method gains less LSD compared with linear models, and the interpolated HRTFs generate perception similar to those of the database.
Introduction
To create immersive 3D audio for virtual reality and gaming, engineers rely on Head-Related Transfer Functions (HRTFs). These functions describe the complex filtering effects of the head, torso, and pinna on sound waves. However, measuring HRTFs for every possible spatial coordinate is impractical. Traditional interpolation methods often rely on linear models like Principal Component Analysis (PCA) or Spherical Harmonics. These models frequently struggle with the inherent nonlinearity of spatial audio cues, particularly when extrapolating to unmeasured distances or elevations.
Methodology and Domain Knowledge
The proposed approach addresses these limitations by integrating physical and perceptual "domain knowledge" into a nonlinear deep learning framework. This integration occurs at two critical stages of the pipeline.
First, instead of using raw spatial coordinates, the model utilizes the Spherical Fourier-Bessel (SFB) basis for feature generation. This basis is derived from the fundamental sound wave propagation equation, providing a mathematically rigorous representation of angular and radial components. Second, the training process is guided by a Log-Spectral Distortion (LSD) loss function. By optimizing for LSD—a metric closely tied to human auditory perception—the model ensures that its predictions are not just mathematically accurate but also perceptually realistic.
Bottleneck DNN Architecture
The core of the nonlinear mapping is a Bottleneck Deep Neural Network (DNN). Unlike standard dense networks, the bottleneck architecture includes a layer with a significantly reduced number of neurons. This forces the network to compress the physics-based SFB features into a compact, high-level representation before reconstructing the HRTF. This compression helps the model ignore noise and capture the essential underlying structures of the spatial audio data.
Key Results
Objective Performance
The Bottleneck DNN significantly outperformed linear SFB models in objective tests. Specifically, the nonlinear model achieved lower LSD across various datasets. In challenging extrapolation scenarios—where the model predicts HRTFs for distances outside the training set—the proposed method showed improvements of up to 4.735 dB. While linear models often fail at low-elevation extrapolation due to body interference, the DNN demonstrated superior robustness.
Subjective Validation
Listening tests involving five subjects confirmed the model's efficacy. Subjects were asked to localize sounds using both measured and model-generated HRTFs. The results indicated that the interpolated HRTFs provided localization accuracy nearly indistinguishable from the original database measurements, validating the perceptual relevance of the LSD-based training.
Conclusion
By combining the predictive power of nonlinear deep learning with the structural guidance of acoustic physics, this research provides a more accurate and robust method for HRTF modeling. Future work will focus on individualization, aiming to adapt this nonlinear model to a user's unique ear anatomy using minimal measurements, such as photographs, to further reduce front-back confusion and improve vertical localization.
Summary updated