Representation of HRTF Based on Common-Pole/Zero Modeling and Principal Component Analysis

Wei Chen, Xiaogang Wei, Hongxu Zhang, Wenpeng HeJournal on Artificial Intelligence 2024 · DOI

Abstract

The Head-Related Transfer Function (HRTF) is essential for rendering accurate 3D audio in immersive multimedia applications, capturing the complex auditory cues used for sound localization. However, HRTF datasets are high-dimensional and nonlinear, requiring significant storage. This study proposes a novel compression and reconstruction method combining a multi-subject common-pole/zero model with Principal Component Analysis (PCA). By estimating common poles across multiple subjects and compressing the resulting zero coefficients via PCA, the researchers achieved a 9.5% compression ratio with an average spectral distortion of less than 2 dB on the CIPIC database.

Introduction

In the realm of Virtual Reality (VR) and immersive audio, the Head-Related Transfer Function (HRTF) acts as a digital fingerprint of how an individual perceives sound from different spatial locations. It accounts for the reflections and scattering caused by the listener's head, torso, and pinnae. While vital for realism, these datasets are massive—often comprising hundreds of thousands of samples. For mobile and consumer-grade hardware, this high dimensionality presents a significant storage and processing bottleneck. Existing models like the Common-Acoustical-Pole/Zero (CAPZ) help by separating direction-independent features (poles) from direction-dependent ones (zeros), but they are typically designed for single individuals and lack sufficient compression for broad practical use.

Methodology

The proposed approach builds upon the traditional CAPZ model. The core innovation lies in extending this model to work across multiple subjects and integrating PCA for enhanced compression. The pipeline involves three main stages:

  1. Multi-Subject Pole Estimation: The human ear canal is modeled as a resonant system. Since its physical properties are relatively stable regardless of sound direction, the researchers estimate "common poles" shared across all directions. By averaging these coefficients across multiple subjects, they create a generalized foundation for the model.
  2. Zero Extraction: With the common poles established, the "zeros" are calculated. These parameters capture the spatial variation and individual anatomical nuances that differentiate one listener's experience from another.
  3. PCA Compression: Principal Component Analysis is applied to the zero coefficients. By identifying the primary directions of variance, the system can discard less significant data while maintaining the overall spectral shape.

Results and Discussion

The method was validated using the CIPIC HRTF database, which contains high-precision measurements for 45 subjects across 1,250 spatial orientations. Performance was measured using Spectral Distortion (SD), where values below 2 dB indicate high perceptual fidelity.

Key findings include:

  • High Efficiency: The system achieved a 9.5% compression ratio while maintaining an average SD of 1.78 dB.
  • Superior Performance: The hybrid model outperformed standard PCA-only approaches. For example, at a 5.5% compression rate, the proposed model reached 4.57 dB distortion, whereas standard PCA required a higher data rate (8.33%) to achieve similar results (4.7 dB).
  • Variance Capture: The study found that a small number of principal components (approximately 14) could capture 99% of the data variance, indicating that the spatial information is highly concentrated within a few dimensions of the coefficient space.

Conclusion

By modeling HRTF coefficients rather than raw spectral data, this research provides a pathway for high-fidelity 3D audio on resource-constrained devices. While the current study focuses on spectral features, future work will address Interaural Time Differences (ITD) and explore linking compressed coefficients to physical anatomical measurements. This could eventually allow for the generation of personalized HRTFs from simple user data, such as photographs of the ear.

Summary updated

Abstract

The Head-Related Transfer Function (HRTF) describes the effects of sound reflection and scattering caused by the environment and the human body when sound signals are transmitted from a source to the human ear. It contains a significant amount of auditory cue information used for sound localization. Consequently, HRTF renders 3D audio accurately in numerous immersive multimedia applications. Because HRTF is high-dimensional, complex, and nonlinear, it is a relatively large and intricate dataset, typically consisting of hundreds of thousands of samples. Storing HRTF requires a significant amount of storage space in practical applications. Based on this, high-dimensional, complex, and nonlinear HRTFs need to be compressed and reconstructed. In this study, inspired by the conventional common-pole/zero model, we propose a method for representing HRTF based on the common-pole/zero model and principal component analysis (PCA). Our method utilizes human auditory features and extends the traditional Common-Acoustical-Pole/Zero (CAPZ) method to estimate the common pole and zero coefficients across multiple subjects. Subsequently, the zero coefficients are compressed using the PCA procedure. Experimental results on the CIPIC database show that the compression ratio can reach 9.5% when the average spectral distortion is less than 2 dB.
BibTeX
@article{chen2024representation,
  author = {Wei Chen and Xiaogang Wei and Hongxu Zhang and Wenpeng He},
  title = {{Representation of HRTF Based on Common-Pole/Zero Modeling and Principal Component Analysis}},
  journal = {Journal on Artificial Intelligence},
  year = {2024},
  doi = {10.32604/jai.2024.052366},
}
Figure