In mobile edge computing (MEC), edge nodes have processing capabilities closer to terminal devices and provide computational services to users at the network edge. However, real-time audio services for mobile users generate computationally intensive mixed audio tasks that require audio separation as a prerequisite, posing challenges for optimizing audio service separation quality under resource-constrained edge environments. In this paper, we propose a lightweight deep learning algorithm for the separation of audio services. Our approach integrates complex ideal ratio mask methods and short-time Fourier transform while incorporating an improved wide inspection block into a streamlined U-Net architecture to enhance audio separation. The objective is to optimize the signal-to-distortion ratio (SDR) of multichannel audio separation under resource limitations while reducing the complexity of the separation model. Through simulation experiments, we demonstrate that our proposed algorithm outperforms conventional approaches in audio service separation, achieving a comparable SDR with lower model overhead. The experimental results highlight the potential of our approach in improving audio service separation in MEC systems.