Gaussian Process Models for HRTF based Sound-Source Localization and Active-Learning

Yuancheng Luo, Dmitry N. Zotkin, Ramani DuraiswamiarXiv 2015 · arXiv

Abstract

Personalizing 3D audio remains a significant challenge due to the unique anatomical cues captured in Head-Related Transfer Functions (HRTFs). This paper proposes a Gaussian Process (GP) framework to address sound-source localization (SSL) and HRTF personalization. By treating SSL as a non-parametric regression problem and employing active learning, the system achieves high localization accuracy with minimal measurements, significantly improving the individual listener's experience.

Introduction

Human sound localization relies on subtle spectral cues shaped by the head, torso, and pinnae, mathematically represented as the Head-Related Transfer Function (HRTF). Two primary hurdles exist in spatial audio: the mapping problem (translating binaural features to 3D directions) and the personalization problem (adapting generic HRTFs to individual listeners). Traditional methods often fail to capture the non-linear acoustic complexities or require exhaustive, impractical measurements in anechoic chambers.

System Architecture and Methodology

The core of the proposed solution is Gaussian Process Regression (GPR). Unlike standard regression, GPR provides both a predictive mean and a variance (uncertainty), which is vital for decision-making in active learning.

System architecture overview showing the integration of GP, Greedy Forward Selection, and Active Learning.

The framework consists of several integrated components:

  • Binaural Feature Extraction: Input features such as Log-magnitude ratios (LMR), Phase differences (PD), and Magnitude pairs (MP) are used to represent sound-source invariant cues.
  • Greedy Forward Selection (GFS): To manage the computational complexity of GPs, GFS identifies a sparse subset of informative HRTF samples, minimizing risk function criteria.
  • Active Learning Recommender: This component uses the GP's uncertainty to select the next HRTF for testing, balancing exploration of unknown spatial regions with exploitation of known high-accuracy matches.

Feature Complexity and Insights

A critical finding of this research is the varying complexity of binaural features. Eigenanalysis of the GP Gram matrices reveals how much information is contained within the leading eigenvalues for different feature types.

Eigenanalysis of GP Gram Matrices for various binaural features (LMR, PD, MP, AMR).

As shown in the analysis, features like Magnitude Pairs (MP) and Phase Differences (PD) can be effectively represented by a small number of leading eigenvalues, suggesting they are smoother and easier to model. Conversely, Log-magnitude ratios (LMR) exhibit higher complexity, requiring more data points for accurate reconstruction. The study also determined that the Squared Exponential kernel is most suitable for modeling the sound-pressure field due to the inherent spatial smoothness of HRTFs.

Results and Discussion

The GP-based approach demonstrates significant improvements in both efficiency and accuracy using the CIPIC HRTF database:

  • Efficiency: High localization accuracy (within 5°) is achievable using only 50 to 150 samples, a massive reduction from the full 1,250-direction grid.
  • Accuracy: GP-SSL models consistently outperformed traditional Ordinary Least Squares (OLS) and Nearest-Neighbor (NN) baselines across all tested binaural features.
  • Personalization: In human-centric simulations, the active-learning framework reduced localization errors by approximately 49% compared to generic HRTFs, effectively "tuning" the audio to the listener's anatomy through minimal interaction.

Conclusion

While the GP framework offers a robust path toward personalized 3D audio, challenges remain regarding computational scaling for real-time applications and modeling human response noise. Future work will focus on extending these models to dynamic, reverberant environments beyond the anechoic conditions studied here.

Summary updated

Abstract

From a machine learning perspective, the human ability localize sounds can be modeled as a non-parametric and non-linear regression problem between binaural spectral features of sound received at the ears (input) and their sound-source directions (output). The input features can be summarized in terms of the individual's head-related transfer functions (HRTFs) which measure the spectral response between the listener's eardrum and an external point in $3$D. Based on these viewpoints, two related problems are considered: how can one achieve an optimal sampling of measurements for training sound-source localization (SSL) models, and how can SSL models be used to infer the subject's HRTFs in listening tests. First, we develop a class of binaural SSL models based on Gaussian process regression and solve a \emph{forward selection} problem that finds a subset of input-output samples that best generalize to all SSL directions. Second, we use an \emph{active-learning} approach that updates an online SSL model for inferring the subject's SSL errors via headphones and a graphical user interface. Experiments show that only a small fraction of HRTFs are required for $5^{\circ}$ localization accuracy and that the learned HRTFs are localized closer to their intended directions than non-individualized HRTFs.
BibTeX
@misc{luo2015gaussian,
  author = {Yuancheng Luo and Dmitry N. Zotkin and Ramani Duraiswami},
  title = {{Gaussian Process Models for HRTF based Sound-Source Localization and Active-Learning}},
  year = {2015},
  eprint = {1502.03163},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1502.03163},
}
Figure