Research background and aims It is well known that real EEGs, as any other electrophysiological activity, are measured relatively to a non-null reference signal. Therefore, the electrical activity recorded by the reference electrode influences all the other measures and might disturb the interpretation of the signals, as well as their automatic analysis (synchronicity evaluation for example). The goal of this communication is to correct the recorded EEGs by estimating the unknown reference signal. Two recording models are possible, depending on the position of the reference electrode, far or close to the recording site. This communication focuses on model I (non-null reference independent from the interest signals), although model II will be briefly exposed. Methods A similar problem was addressed by Hu (2007), the proposed solution being based on an ICA algorithm. We propose here a new method, faster and more accurate, also based on an ICA model. During a first step, a normalized reference is estimated using the particular structure of the mixing matrix in the case of a non-null referenced model. Second, the amplitude of the reference is estimated by minimizing a mean squared error (MSE) type criterion. Results and conclusions Several simulations are used to illustrate the proposed method. Using the MSE between the estimated and the simulated reference signal, we show that our method outperforms previously proposed methods for all noise levels. Coherence estimations for the simulated signals are also presented, before and after reference correction. An example on a real depth EEG signal concludes the communication.