New SFF-based DoA estimator outperforms GCC methods in noisy, reverberant settings

Direction-of-arrival (DoA) estimation from microphone signals remains difficult in noisy, reverberant, multi-speaker environments. Conventional approaches like generalized cross-correlation (GCC) operate in the short-time Fourier transform (STFT) domain, which primarily reflects vocal-tract characteristics. However, these spectral features degrade under adverse conditions. Now, researchers propose a new method based on single frequency filtering (SFF) that provides high spectral resolution of harmonics alongside high temporal resolution of excitation-source events (like epoch impulses). By correlating the envelopes of SFF outputs across microphone channels using PHAT-weighted GCC, the method leverages source features that are inherently more robust to noise and reverberation than spectral features.

The team evaluated their SFF-based DoA estimator against state-of-the-art GCC-based estimators using publicly available real-room recordings under challenging conditions: high reverberation, multiple simultaneous speakers, and additive noise. Results show the proposed method—and an existing SFF-based estimator—delivers detection and accuracy performance that is superior or comparable to the best GCC-based estimator across all test cases. Additionally, using only speech-dominant frequency bins significantly improves GCC-PHAT robustness, suggesting a straightforward improvement for future SFF-based DoA systems. This work opens a path toward more reliable acoustic source localization for hands-free communication, hearing aids, and smart audio devices.