視聽障人士在日常生活中難以感知環境聲音的方向與內容，現有單麥克風輔助裝置無法提供空間聲音資訊。本研究整合四麥克風陣列與SEMamba語音增強模型，建構適用於資源受限環境的多通道音訊處理系統，系統性比較廣義旁瓣消除器（Generalized Sidelobe Canceller， GSC）與延遲加總（Delay-and-Sum， DAS）波束形成方法在不同噪音環境下的性能差異。研究採用合成音訊與實際錄音雙軌實驗，合成音訊實驗結果顯示，在在訊噪比（Signal-to-Noise Ratio， SNR）為-8 dB、μ=1.0條件下，GSC波束形成器相較單聲道系統，在雜訊0°方向可將字元錯誤率由0.877降至約0.046（降低約94.7%），在90°方向由0.654降至約0.019（降低約97.1%），相較於DAS固定波束形成，在SNR ≤ -10 dB的極低訊噪比條件下，DAS在所有四個方向均完全失敗（Character Error Rate， CER=1.0），而GSC在0°方向可將CER降至約0.26，在90°、180°、270°方向可降至0.035至0.050，均能維持部分辨識能力。步進係數μ在低訊噪比環境下，介於0.5至1.0的設定能取得較佳的性能平衡。SEMamba語音增強模型與波束形成器在合成音訊與實際錄音實驗中均展現整體正面效果，兩者在低訊噪比區間的效果更加顯著；SEMamba的輔助效果高度依賴基礎方法是否仍有錯誤可供改善。本研究驗證了整合波束形成與深度學習語音增強之多通道輔助系統架構的可行性，可作為未來嵌入式輔助裝置設計的研究基礎。;Individuals with visual and hearing impairments face difficulties in perceiving the direction and content of environmental sounds. Existing assistive devices predominantly rely on single-microphone systems， unable to provide spatial sound information. This study integrates a 4-microphone array with the SEMamba speech enhancement model to construct a multi-channel audio processing system for resource-constrained environments， systematically comparing Generalized Sidelobe Canceller (GSC) and Delay-and-Sum (DAS) beamforming methods under various noise conditions. The research employs synthetic audio and real recording experiments. In the synthetic experiments， at Signal-to-Noise Ratio (SNR) = −8 dB with μ = 1.0， GSC reduces CER from 0.877 to approximately 0.046 (94.7% reduction) at the 0° noise direction， and from 0.654 to approximately 0.019 (97.1% reduction) at 90°. Compared to DAS fixed beamforming， at SNR ≤ −10 dB， DAS completely fails in all four directions (Character Error Rate， CER = 1.0)， while GSC reduces CER to approximately 0.26 at 0° and to 0.035–0.050 at 90°， 180°， and 270°， maintaining partial recognition capability. Step size coefficient μ in the range of 0.5 to 1.0 provides the best performance balance under low SNR conditions. Both synthetic and real recording experiments demonstrate overall positive effects of SEMamba， with more pronounced improvements in the low SNR range; the degree of improvement is highly dependent on whether the baseline method still has residual errors to correct. This study validates the feasibility of a multi-channel assistive system architecture integrating beamforming with deep learning speech enhancement， providing a research foundation for future embedded assistive device design.
