Sparse Head-Related Impulse Response for Efficient Direct Convolution

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

Abstract

Head-related impulse responses (HRIRs) are subject-dependent and direction-dependent filters used in spatial audio synthesis. They describe the scattering response of the head, torso, and pinnae of the subject. We propose a structural factorization of the HRIRs into a product of non-negative and Toeplitz matrices; the factorization is based on a novel extension of a non-negative matrix factorization algorithm. As a result, the HRIR becomes expressible as a convolution between a direction-independent resonance filter and a direction-dependent reflection filter. Further, the reflection filter can be made sparse with minimal HRIR distortion. The described factorization is shown to be applicable to the arbitrary source signal case and allows one to employ time-domain convolution at a computational cost lower than using convolution in the frequency domain.

Introduction

To create realistic 3D audio in headphones, engineers use Head-Related Transfer Functions (HRTFs)—or their time-domain counterparts, HRIRs—to simulate how sound interacts with a listener's anatomy. A significant challenge in real-time spatial audio is the computational cost of convolving high-fidelity audio streams with these long filters. While frequency-domain methods like the Fast Fourier Transform (FFT) are commonly used for efficiency, they introduce block-processing latency that can be problematic for interactive applications like virtual reality. This paper proposes a sparse representation of HRIRs that enables ultra-fast, low-latency convolution directly in the time domain.

Proposed Methodology

The core innovation is a structural factorization of the HRIR ($x$) into a convolution of two simpler components: a direction-independent resonance filter ($f$) and a direction-dependent reflection filter ($g$), such that $x \approx f * g$.

Structural factorization of HRIR into resonance and reflection components.
  1. The Resonance Filter ($f$): Represents shared acoustic characteristics, such as ear canal resonances, for a specific subject across all directions.
  2. The Reflection Filter ($g$): Captures discrete reflections off the anatomy. This component is constrained to be non-negative and sparse to align with physical sound propagation.

To achieve this, the authors developed a Modified Semi-Non-negative Matrix Factorization (semi-NMF) algorithm. By introducing a Toeplitz constraint on the basis matrix, the factorization is forced to behave as a convolution. Sparsity is further refined using $L_1$-regularized non-negative least squares ($L_1$-NNLS), allowing for a precise trade-off between the number of non-zero coefficients and reconstruction error.

Experimental Setup

The approach was evaluated using the CIPIC HRTF database, focusing on high-resolution HRIR measurements. Data was pre-processed using minimum-phase conversion and initial delay removal. Subject 003 from the database was used for detailed performance analysis.

Results and Discussion

The algorithm demonstrates a clear trade-off between filter sparsity and reconstruction accuracy. As the regularization term $\lambda$ increases, the number of non-zero entries in the reflection filter decreases, leading to a gradual rise in Root Mean Square Error (RMSE) and Spectral Distortion (SD).

Trade-off between sparsity (non-zero entries) and reconstruction error (RMSE, SD) for a sample HRIR.

Key results include:

  • Sparsity vs. Error: The algorithm achieved a mean SD of 3.0 dB with a full reflection filter and 5.3 dB when sparsified to roughly half its original size.
  • Computational Crossover: Theoretical calculations suggest the sparse time-domain approach is faster than state-of-the-art frequency-domain implementations when the number of non-zero elements is less than ~127 for long signals, or ~90 for shorter segments typical of low-latency buffers.
  • Spatial Performance: The model is particularly effective for ipsilateral directions (the ear facing the source). Contralateral directions present more complex acoustic responses that are harder to sparsify with short filters.
  • Frequency Weighting: Utilizing a Gaussian windowing transform during optimization prioritized low-frequency accuracy, improving spectral distortion by approximately 10%.

Conclusion and Future Work

This work provides a framework for efficient spatial audio by factorizing HRIRs into resonance and sparse reflection components. This method allows time-domain convolution to become more computationally efficient than FFT-based methods in low-latency scenarios. Future research will focus on improving contralateral modeling, potentially through spatially-localized sub-models, and evaluating performance across diverse hardware architectures.

Summary updated

Abstract

Head-related impulse responses (HRIRs) are subject-dependent and direction-dependent filters used in spatial audio synthesis. They describe the scattering response of the head, torso, and pinnae of the subject. We propose a structural factorization of the HRIRs into a product of non-negative and Toeplitz matrices; the factorization is based on a novel extension of a non-negative matrix factorization algorithm. As a result, the HRIR becomes expressible as a convolution between a direction-independent \emph{resonance} filter and a direction-dependent \emph{reflection} filter. Further, the reflection filter can be made \emph{sparse} with minimal HRIR distortion. The described factorization is shown to be applicable to the arbitrary source signal case and allows one to employ time-domain convolution at a computational cost lower than using convolution in the frequency domain.
BibTeX
@misc{luo2015sparse,
  author = {Yuancheng Luo and Dmitry N. Zotkin and Ramani Duraiswami},
  title = {{Sparse Head-Related Impulse Response for Efficient Direct Convolution}},
  year = {2015},
  eprint = {1502.03162},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1502.03162},
}

Loading related content...

Figure