Addressing inefficiencies in blind source separation and limitations in dynamic signal disentanglement within complex audio environments, this study introduces a novel approach based on Neurally Plausible Alternating Optimization-based Online Dictionary Learning (NOODL). This method overcomes the shortcomings of traditional dictionary learning techniques, such as slow processing speeds and poor adaptability to dynamic signals. The NOODL algorithm enhances the dictionary's sparse representation capability through a highly flexible and robust updating mechanism, making it especially suitable for processing complex speech signals. Experiments conducted in a MATLAB environment using the UrbanSound8K audio dataset demonstrate the NOODL algorithm's significant impact on improving signal recovery accuracy and separation quality. Compared to traditional methods, NOODL shows superior performance by increasing signal correlation coefficients, reducing mean square errors, and enhancing both the signal-to-noise ratio and peak signal-to-noise ratio.