[Paper Review] On the Impact of Localization Errors on HRTF-based Robust Least-Squares Beamforming

Hendrik Barfuss, Walter Kellermann|arXiv (Cornell University)|Mar 29, 2016

Speech and Audio Processing9 references3 citations

TL;DR

This paper investigates the impact of localization errors on HRTF-based robust least-squares beamforming in robot audition systems. It demonstrates that errors in direction-of-arrival (DOA) estimation or robot-source distance mismatch degrade beamforming performance, especially when the beamformer is steered toward interferers, but HRTF-based designs still outperform free-field alternatives even under such errors.

ABSTRACT

In this work, a recently proposed Head-Related Transfer Function (HRTF)-based Robust Least-Squares Frequency-Invariant (RLSFI) beamformer design is analyzed with respect to its robustness against localization errors, which lead to a mismatch between the HRTFs corresponding to the actual target source position and the HRTFs which have been used for the beamformer design. The impact of this mismatch on the performance of the HRTF-based RLSFI beamformer is evaluated, including a comparison to the free-field-based beamformer design, using signal-based measures and word error rates for an off-the-shelf speech recognizer.

Motivation & Objective

• To evaluate the robustness of HRTF-based RLSFI beamformers under localization errors in robot audition scenarios.
• To investigate the performance degradation caused by incorrect DOA estimates (azimuth and elevation) in HRTF-based beamformer design.
• To assess the impact of robot-source distance mismatch on beamformer performance when HRTFs are measured at a different distance than actual source position.
• To compare HRTF-based beamforming with conventional free-field-based beamforming under localization errors.
• To provide design guidelines for HRTF-based beamformers by quantifying sensitivity to localization inaccuracies.

Proposed method

• Adapts the robust least-squares frequency-invariant (RLSFI) beamformer design to use measured or simulated HRTFs instead of free-field steering vectors.
• Replaces free-field sensor responses in the LS optimization with HRTF-based responses, where gn,HRTF(ωp, φm, θm) = hmn(ωp), modeling direct path propagation.
• Imposes constraints on white noise gain (WNG ≥ γ) and distortionless response in the desired look direction (wT_f(ωp)d(ωp) = 1).
• Solves the LS optimization problem for each discrete frequency ωp, then performs FIR approximation of the optimal filter coefficients.
• Uses HRTFs measured at a reference robot-source distance (1.1m) to design the beamformer, while testing performance at different distances (e.g., 2m).
• Evaluates performance using ASR metrics: word error rate (WER) and frequency-weighted segmental SNR (fwSegSNR), averaged over multiple source positions.

Experimental results

Research questions

• RQ1 How do azimuthal DOA estimation errors affect the performance of HRTF-based RLSFI beamformers?
• RQ2 What is the impact of robot-source distance mismatch on beamformer output quality when HRTFs are designed for a different distance?
• RQ3 How does the beamformer performance degrade when the beam is steered toward an interferer due to localization errors?
• RQ4 Does the HRTF-based beamformer maintain performance advantages over free-field-based beamformers under localization errors?
• RQ5 Can spatial nulls of the beampattern be exploited to mitigate interference when localization errors shift the beam away from the interferer?

Key findings

• Localization errors of ±5° and ±10° in azimuth degrade HRTF-based beamformer performance, increasing average WER from 31.4% to 40.5% when the beam is steered toward the interferer.
• When localization errors steer the beam away from the interferer (e.g., +5° or +10°), performance improves due to the beam pattern’s spatial null approaching the interferer’s direction.
• A robot-source distance mismatch of 1.1m (design) vs. 2m (actual) causes a 13.5° elevation mismatch, resulting in a 9.1% to 8.8% increase in WER and a slight decrease in fwSegSNR.
• Despite degradation under errors, the HRTF-based beamformer consistently outperforms the free-field-based beamformer in both WER and fwSegSNR across all tested conditions.
• The performance degradation is most severe when the beamformer is steered toward the interferer, while errors that move the beam away from interference can improve suppression.
• The study confirms that accurate localization is critical for HRTF-based beamformer design, as mismatched HRTFs significantly reduce robustness and output quality.