Due to the high cost of modeling, model-based techniques are yet to make their impact in the embedded systems industry, which still persist on maintaining code-oriented legacy systems. Re-engineering existing code-oriented systems to fit model-based development is a risky endeavor due to the cost and efforts required to maintain correspondence between the code and model. We aim to reduce the cost of modeling and model maintenance by automating the process, thus facilitating model-based techniques. We have previously proposed the use of automatic model extraction from recordings of existing embedded real-time systems. To estimate the quality of the extracted models of timing behavior, we need a framework for objective evaluation. In this paper, we present such a framework to empirically test and compare extracted models, and hence obtain an implicit evaluation of methods for automatic model extraction. We present a set of synthetic benchmarks to be used as test cases for emulating timing behaviors of diverse systems with varying architectural styles, and extract automatic models out of them. We discuss the difficulties in comparing response time distributions, and present an intuitive and novel approach along with associated algorithms for performing such a comparison. Using our empirical framework, and the comparison algorithms, one could objectively determine the correspondence between the model and the system being modeled.