Ego-Noise-Aware Spatial Filtering for Reliable UAV Audition in Extreme Low-SNR Conditions

TL;DR — This paper presents a lightweight, model-free speech enhancement framework tailored for UAV audition systems operating under extreme ego-noise, achieving 98.12% DoA accuracy at -25 dB SNR. It uses an ego-noise-aware spatial filtering pipeline that jointly optimizes DoA estimation, crossover-based hybrid noise covariance matrix estimation, and a variance-modulated post-filter.

Key contributions

• Phase-consistency-guided time-frequency bin selection for robust direction-of-arrival (DoA) estimation that mitigates false alarms from temporally correlated drone noise.
• A crossover-based hybrid noise covariance matrix (NCM) estimation strategy that combines closeness gains at low frequencies and directional gains at high frequencies.
• A null-reference post-filter utilizing variance modulation to suppress residual stationary ego-noise without distorting the target speech.
• An ultra-lightweight signal processing architecture requiring only 0.021 GMAC/s without relying on pre-trained neural networks.

Problem

Unmanned aerial vehicle (UAV) audition systems struggle under severe ego-noise characterized by strong temporal correlation, frequency-varying composition, and amplitude stationarity. Traditional spatial filtering methods fail because these noise properties destabilize direction-of-arrival (DoA) estimation, corrupt noise covariance matrix (NCM) estimation, and leave persistent residual noise after beamforming. Prior deep learning solutions demand heavy computation exceeding several GMAC/s, restricting real-time on-device deployment.

Method

The system processes multichannel audio captured by a circular microphone array (radius 0.1 m, 4 mics) via the STFT domain (512 sample window, 75% overlap). First, a phase-consistency-guided temporal similarity metric computes phase delays between microphone pairs, and a dynamic threshold based on Median Absolute Deviation (MAD with k=2.5) selects speech-informative TF bins. Spatial likelihood is then aggregated over these selected bins to determine the global DoA via statistical mode. This shared DoA reference drives subsequent spatial filtering stages.

Next, a frequency-adaptive hybrid masking strategy is used to estimate the signal covariance matrix (SCM) and NCM for an MVDR beamformer. A crossover frequency nu is defined at 3 kHz (separating regimes based on the Speech Intelligibility Index). Below nu, a closeness gain (assuming a 10-degree DoA standard deviation) handles lower frequencies where directional resolution is poor; above nu, a directional gain combining maximum directivity factor (mDF) and delay-and-sum (DS) beamformers takes over to overcome high-frequency ego-noise dominance.

Finally, a null-reference post-filter suppresses stationary residual ego-noise from the MVDR output. Using a null direction steering vector corresponding to the minimum speech likelihood over selected bins, a null-space projection estimates residual noise. A variance-modulated primary Wiener gain scales suppression depending on signal versus noise variance, driving the gain toward unity when speech is present to prevent signal distortion.

Experimental setup

Evaluated on 160 TIMIT speech sentences synthesized with real Syma Z4W drone ego-noise via Pyroomacoustics under both anechoic and reverberant (T60 = 0.3 s) conditions. Target DoAs were set to {-45, 0, 45} degrees across five input SNR levels (-25, -20, -15, -10, -5 dB), yielding 480 samples per SNR. Baselines include MPDR, direction-gain MVDR, closeness-gain MVDR, and MMSE beamformers. Metrics include SI-SDR, PESQ, ESTOI, DNSMOS P.835 (SIG, BAK, OVRL), DoA accuracy, and Mean Absolute Error (MAE).

Results

At an input SNR of -25 dB under anechoic conditions, the proposed method achieves an SI-SDR of -2.446 dB (outperforming the best baseline MPDR variant at -10.150 dB), a PESQ of 1.449, an ESTOI of 0.249, and a DNSMOS OVRL of 2.245. Under reverberant conditions at -25 dB SNR, it reaches an SI-SDR of -5.134 dB and a PESQ of 1.379. DoA estimation accuracy reaches 98.12% in anechoic and 94.79% in reverberant conditions at -25 dB SNR. Ablations show that adding the proposed post-filter consistently improves PESQ and background noise (BAK) scores without harming speech intelligibility.

Limitations

The evaluation is restricted to a simulated setup using static hovering drone recordings and synthetic room impulse responses. The framework assumes a stationary UAV state and does not account for acoustic dynamics or aerodynamic turbulence shifts during active flight or non-hovering movement maneuvers. Generalization to variable multi-drone environments or uncalibrated array geometries remains unproven.

Why read this

Researchers and engineers designing real-time, resource-constrained audio front-ends for robotic platforms will learn how to tightly couple DoA estimation and hybrid spatial covariance masking using domain-specific noise properties rather than heavy neural networks.

Code

None released (as of this page's updated date). If you are an author with a repo, please claim this entry — see CONTRIBUTING.md.

Applications

On-board speech recognition, emergency communication, and acoustic human-robot interaction for unmanned aerial vehicles in noisy outdoor environments.

Institutions

Sungkyunkwan University, University of Illinois Urbana-Champaign

Funding / 經費: National Research Foundation, IITP, KIAT

Related

• DroFiT: A Lightweight Band-Fused Frequency Attention Toward Real-Time UAV Speech Enhancement — same problem · relatedness 2.5/3
• Joint Learning of Covariance Estimation and White Noise Gain for Robust MVDR Beamforming — same problem · relatedness 2.0/3
• Fast Multichannel Nonnegative Matrix Factorization with Directivity Regularization for DOA-Informed Speech Separation — same problem · relatedness 2.0/3
• SpkGuideDOA: Speaker-wise Representation Guidance for Multiple Moving Speaker Localization — same problem · relatedness 2.0/3
• Towards Array-Invariant Speech Enhancement via Geometry-Aware Dynamic Convolution — same problem · relatedness 2.0/3

All 950k paper pairs scored by TypeSafe Jev (scripts/related/); relatedness 0 = unrelated … 3 = directly comparable.