This paper presents an algorithm to perform online tremor characterization from motion sensors measurements, while filtering the voluntary motion performed by the patient. In order to estimate simultaneously both nonstationary signals in a stochastic filtering framework, pathological tremor was represented by a time-varying harmonic model and voluntary motion was modeled as an auto-regressive moving-average (ARMA) model. Since it is a nonlinear problem, an extended Kalman filter (EKF) was used. The developed solution was evaluated with simulated signals and experimental data from patients with different pathologies. Also, the results were comprehensively compared with alternative techniques proposed in the literature, evidencing the better performance of the proposed method. The algorithm presented in this paper may be an important tool in the design of active tremor compensation systems.