User-specific subspace method in speaker recognition, such as CLAFIC is used to construct the individual subspace which represents the distinct spectral characteristic for each speaker. It concatenates the speech matrices side by side before forming the correlation matrix. In this paper, we proposed a new method, coined as Probabilistic 2D CLAFIC which applied a straightforward two-dimensional (2D) speech matrix for feature dimension reduction to improve the discrimination and reduce the computation complexity. Gaussian Mixture Model (GMM) is used instead of norm of the projected feature in conventional CLAFIC to boost up the performance. Experimental results showed that our proposed method attained an encouraging performance with the best Equal Error Rate (EER) of 0.56% with full feature dimension, and EER of 2.98% with 50% reduction of feature dimension compare to baseline GMM, 5.36%.