<p>Speech separation aims to separate target speech from overlapping mixed speech signals. Previous approaches in speech separation have limited capability in modeling long sequence audio, while suffering from high computational complexity. Recent approaches highlight the importance of multi-scale features for audio modeling but are limited by the homogeneous semantic content in long-sequence audio. To address these issues, we propose a speech separation network based on an improved state space model (SS-SSM) termed MSMa-Net. Specifically, we decompose the input mixture into multi-scale representations at different resolutions, encoding rich multi-scale acoustic features. Then, the SS-SSM is employed to reduce semantic redundancy in multi-scale features, enhancing the model’s capability for long-term audio modeling. Experiments show that MSMa-Net matches state-of-the-art methods while reducing computational complexity and latency, making it a promising solution for practical applications.</p>