Proportional variance dependency among the frequency components is characteristic of natural signals and has been utilized in frequency-domain blind source separation to solve the permutation problem. In order to increase robustness in such methods, overall measures have been preferred to the measures between directly neighboring frequency components. The overall variance dependency pattern in the fullband, however, can vary by signals and is difficult to be modeled, whereas in smaller subbands the proportional variance dependency is more definite. Here, a novel permutation correction method that utilizes the proportional variance dependency in small subbands is proposed. A windowed likelihood function that uses source priors with internal variance dependency is employed as the measure of permutation correction. This method not only shows robust separation performance but also is computation-wise very efficient.