In this article we propose a method to improve the accuracy of sinusoid modeling by introducing parameter variation models into both the analyzer and the synthesizer. Using the least-square-error estimator as an example, we show how the sinusoidal parameters estimated under a stationary assumption relate to the real nonstationary process, and propose a way to reestimate the parameters using some parameter variation model. For the synthesizer, we interpolate the parameters using the same model, with the phase unwrapping process reformulated to adapt to the change. Results show that the method effectively cuts down the systematic error of a conventional system based on a least-square-error estimator and the McAulay-Quatieri synthesizer.