Dysarthric speech recognition presents significant challenges due to limited resources and the diverse attributes of such speech. Current solutions often involve techniques like data augmentation or domain adaptation. However, applying these methods to a single model frequently results in inconsistent improvements across different speech intelligibility groups. To address this, we propose a novel approach that combines curriculum learning with a multi-stream architecture. Experimental results show the effectiveness of the proposed method, achieving a 10.47% improvement over the baseline.