SFF-based narrowband DoA estimator outperforms broadband methods

Direction-of-arrival (DoA) estimation from distant microphones is critical for applications like smart speakers and hearing aids. Broadband (BB) methods aggregate across all frequencies to avoid spatial aliasing but struggle to resolve multiple simultaneous speakers in a single time frame. Narrowband (NB) methods exploit frequency sparsity to achieve multi-speaker DoA estimation per frame, yet they are vulnerable to spatial aliasing. In a new preprint, Thakallapalli et al. propose a NB estimator that leverages single frequency filtering (SFF) to select time-frequency regions with high signal-to-noise ratio. By cross-correlating these SFF-filtered regions across microphones, the estimator gains robustness against spatial aliasing while preserving the multi-speaker capability of narrowband approaches.

The authors compared their SFF-based NB estimator against four state-of-the-art methods (one NB, three BB) using detection and accuracy metrics on simulated and real-world data under varying reverberation and noise conditions. Results show that the SFF-based approach consistently outperforms the existing NB baseline and surpasses some BB methods. The key innovation lies in SFF's ability to discriminate speech from non-speech even in degraded conditions, providing cleaner cross-correlation peaks. This work promises more accurate and reliable multi-speaker localization in challenging acoustic environments, with potential impact on voice interfaces, teleconferencing, and auditory scene analysis.