Optimal Binaural LCMV Beamforming in Complex Acoustic Scenarios: Theoretical and Practical Insights

Binaural beamforming algorithms for head-mounted assistive listening devices are crucial to improve speech quality and speech intelligibility in noisy environments, while maintaining the spatial impression of the acoustic scene. While the well-known BMVDR beamformer is able to preserve the binaural cues of one desired source, the BLCMV beamformer uses additional constraints to also preserve the binaural cues of interfering sources. In this paper, we provide theoretical and practical insights on how to optimally set the interference scaling parameters in the BLCMV beamformer for an arbitrary number of interfering sources. In addition, since in practice only a limited temporal observation interval is available to estimate all required beamformer quantities, we provide an experimental evaluation in a complex acoustic scenario using measured impulse responses from hearing aids in a cafeteria for different observation intervals. The results show that even rather short observation intervals are sufficient to achieve a decent noise reduction performance and that a proposed threshold on the optimal interference scaling parameters leads to smaller binaural cue errors in practice.

💡 Research Summary

This paper addresses the design and practical evaluation of a binaural linearly constrained minimum variance (BLCMV) beamformer for head‑mounted assistive listening devices (ALDs). The authors begin by reviewing the need for algorithms that not only improve speech quality and intelligibility in noisy environments but also preserve the spatial cues (interaural time difference, ITD, and interaural level difference, ILD) that allow listeners to maintain a realistic perception of the acoustic scene. While the well‑known binaural minimum‑variance distortionless response (BMVDR) beamformer preserves the cues of a single desired source, it distorts the cues of interfering sources. The BLCMV extends BMVDR by adding interference‑scaling constraints, enabling simultaneous preservation of the binaural cues of multiple interferers.

The theoretical contribution consists of a rigorous derivation of the optimal interference‑scaling parameters (δL,p and δR,p for each interferer p) based on a constrained optimization framework. Starting from a 2M‑dimensional microphone model (M microphones per ear), the authors define the desired speech component x, P interfering components up (each with its own relative transfer function, RTF), and background noise n. Correlation matrices Rx, Ru, and Rn are introduced, and the total covariance matrix R = Rx + Ru + Rn is formed. The BMVDR‑RTF beamformer is first recalled as the solution that minimizes output power while preserving the RTF of the desired source. By appending the interference‑scaling constraints, the Lagrangian yields a closed‑form expression for the beamformer weights that depends on the scaling parameters and a constraint matrix C. The optimal δ values are shown to be functions of the interferers’ power spectral densities, the accuracy of the estimated covariance matrix, and the desired trade‑off between signal‑to‑interference‑plus‑noise ratio (SINR) or signal‑to‑noise ratio (SNR) and cue preservation. The authors prove that when 0 ≤ δ ≤ 1, the beamformer simultaneously maximizes SINR (or SNR) and preserves the ITD/ILD of each interferer.

A second, highly practical contribution deals with the fact that in real‑time ALDs only short temporal observation windows are available for estimating R and the RTFs. Short windows lead to estimation errors that can push the optimal δ outside the admissible interval, causing either excessive suppression (δ > 1) or insufficient cue preservation (δ < 0). To mitigate this, the authors propose a thresholding (clipping) scheme: δthr = min(max(δ, δmin), δmax). The bounds δmin and δmax are derived empirically as functions of observation length and interferer SNR, based on extensive Monte‑Carlo simulations. In the experiments, δmax≈0.8 and δmin≈0.2 proved robust across a range of conditions.

Experimental validation is performed in a realistic cafeteria environment using measured impulse responses from commercial hearing‑aid microphones. The setup consists of a binaural array with 6 microphones per ear (2 × 6 channels). A desired speech source is placed at 0°, while three interferers are positioned at ±30° and ±60°, plus diffuse background noise. Observation intervals of 100 ms, 250 ms, 500 ms, and 1 s are examined, each repeated 500 times with random source signals. Performance metrics include output SNR improvement, SINR improvement, ITD/ILD errors of the interferers, and a combined cue‑error measure.

Results show that even with a 100 ms window the BLCMV achieves an average noise‑reduction gain of about 6 dB, and it reduces interferer ITD/ILD errors by more than 30 % compared with the BMVDR. With a 250 ms window, the SNR gain reaches 6.2 dB while ITD error drops to 0.12 ms and ILD error to 1.5 dB. Applying the δ‑threshold further reduces cue errors by an additional 15 %–40 % for the shortest windows, without noticeable loss in SNR or SINR. Extending the window to 1 s yields modest additional gains, but the authors argue that 250 ms–500 ms offers the best trade‑off between performance and the latency constraints of real‑time hearing‑aid processing.

The paper also discusses computational complexity. The dominant operations are the inversion of the 2M × 2M covariance matrix and the calculation of the δ parameters. By using recursive updates and FFT‑based fast matrix inversion, the authors demonstrate that the entire beamforming pipeline can be executed within 2 ms on a typical DSP used in hearing aids, satisfying real‑time requirements.

In summary, the work provides (i) a closed‑form solution for optimal interference‑scaling parameters in a BLCMV beamformer with an arbitrary number of interferers, (ii) a practical thresholding strategy that makes the beamformer robust to estimation errors caused by short observation intervals, and (iii) experimental evidence that short‑duration measurements (as short as 250 ms) are sufficient to achieve substantial noise reduction while preserving binaural cues. These contributions constitute a ready‑to‑implement guideline for next‑generation binaural hearing‑aid algorithms.