This paper considers the mixture models to unify software reliability models (SRMs). The modeling framework is based on the randomization of a software fault detection rate. Our models, called the mixed SRMs in this paper, can include some existing SRMs such as Littlewood NHPP model and the hyperexponential SRM. We develop an efficient parameter estimation algorithm for the mixed SRMs based on the EM (Expectation-Maximization) principle, and investigate both the effectiveness of the EM algorithm and the validity of the mixed SRMs through the numerical examples with real software failure data.