Neural Network-Assisted Joint DOA Estimation and Beamforming with First-Order Reflection Modeling

Source signal extraction and direction-of-arrival (DOA) estimation are core tasks in microphone array signal processing. However, their performance degrades significantly in reverberant environments: DOA estimation struggles to distinguish the direct path from strong reflections, while source extraction fails to effectively exploit the spatial cues in early reflections. This paper proposes a novel formulation of the source array-manifold vector (AMV) as a linear combination of free-field AMVs corresponding to the direct path and the four first-order reflections from the surrounding walls. A neural network is developed to jointly estimate the DOA and the combination weights of these component AMVs. The resulting composite AMV is then used in a minimum variance distortionless response (MVDR) beamformer for enhanced source extraction. Additionally, we introduce a location-based training strategy that enables end-to-end learning to automatically identify the direct-path direction. Simulation results show that the proposed method significantly improves both speech quality and DOA estimation accuracy in reverberant environments.