Introduction: We present an extension of an artificial ECG model for the simulation of multi-lead ECG during paroxysmal Atrial Fibrillation (AF). Method: This new method was based on adaptation of previously published models, one which generates the cardiac dipole from a sum of Gaussian kernels and one which simulates atrial activity during AF by using a sawtooth like shaped function. The proposed model also allows for the generation of ectopic beats. The generation of paroxysmal AF was made possible by the use of a multiple layer-based Hidden Markov Model (HMM). The first layer allowed for the switching between different rhythms (AF or normal sinus rhythm). The second layer enabled the simulation of short and long RR intervals, as AF is usually accompanied with the succession of short and long RR intervals. The values of the transition matrix were defined as described in the literature for realistic AF rhythmicity. The final HMM layer allowed for the simulation of ectopic beats independently of the rhythm. The cardiac dipole was then generated with a sum of Gaussian kernels appropriate to the beat type. During AF, the Gaussians for the P wave were replaced by the saw-tooth like atrial activity function. Each parameter of the different models was allowed to evolve stochastically to simulate the inherent non-stationarity of a physiological signal. Finally, physiological noise extracted from the noise-stress database were added to enhance the realism of the simulation. Results: The model was first visually inspected in order to assess its realism. State-of-the-art RR interval based AF detection and atrial activity techniques were applied on the simulated signals and results were compared to the literature. AF detection results were lower than those reported in the literature (Se=62%), but atrial activity extracted were in accordance with previously published results. Discussion: The simulation of ECG signal with paroxysmal AF provides a framework for quantitative assessment of any atrial activity extraction technique, and is useful for preliminary validation before application on real ECG data, or evaluation on rare events that can be simulated instead.