Robust DOA Estimation: Cross-Branch Fused Multi-stream Network with Feature Enhancement

Accurate direction of arrival (DOA) estimation is critical for many audio signal processing tasks in microphone array applications. However, real-world environments pose considerable challenges to DOA estimation algorithms due to noise and reverberation. Therefore, we propose a robust DOA estimation framework. First, the signal power at the zeroth-degree zeroth-order in spherical harmonic domain is used to identify reliable time–frequency bins which represent an omnidirectional field with no variation in the source directions. Two masks are generated by a proportional threshold and a deep neural network, respectively. They are combined with the spherical harmonic features, resulting in two enhanced complex features. Then, we design a new cross-branch fused multi-stream network (CMnet) to exploit the interaction structure between branches, which greatly digs into the specificity and complementarity of real and imaginary parts of the enhanced complex features. Experimental evaluations on simulated data set and LOCATA data set show that our methods have better DOA estimation performance compared with the existing methods, especially in challenging environments, and they are also applicable to real-world scenarios.