Modern devices such as teleconferencing systems, robots, augmented reality headsets, and smart environments rely on microphone arrays to capture spatial sound scenes. Unlike single microphones, arrays record sound at multiple spatial positions, enabling downstream processing such as source localization, spatial rendering, and immersive audio reproduction. However, background noise remains a major obstacle: it degrades signal quality and disrupts subsequent spatial processing. This project develops a new method for direction-preserving noise reduction for microphone arrays. Instead of producing a single denoised signal, the proposed approach processes all microphone channels jointly and outputs a denoised multichannel signal. This preserves the spatial relationships between microphones, which are essential for applications such as sound source localization, spatial audio reproduction, and human–machine interaction. The core idea is to combine modern neural networks with model-based multichannel signal processing. A neural network will estimate the spatial characteristics of background noise in real time, represented as time-varying covariance matrices. These estimates are then used in a mathematically grounded subspace decomposition framework to separate noise from the spatial sound scene while maintaining geometric consistency across channels. This hybrid design leverages the robustness of data-driven learning and the interpretability and stability of classical signal processing.