A novel method to improve the separation performance and the convergence speed of fast independent vector analysis (FastIVA) for frequency domain operation is proposed. In this paper, a mixture of super Gaussian distributions is adopted as a source prior for the FastIVA algorithm. In this mixed source prior, a Student's t distribution is adopted to model the high amplitude information in the spectrum of a speech signal and another super Gaussian distribution with less heavy tails is used to model the rest of the information. Moreover, in the proposed mixed source prior the weight of both distributions can be varied according to the frequency dependent amplitude of the speech signals. The performance of the proposed algorithm is demonstrated by using the imaging method and binaural room impulse responses and comparison is made with the FastIVA using the original super Gaussian source prior.