This paper is concerned with the identification of the modal parameters of structures. It is a well known result that, due to the presence of measurement noise, time domain identification methods often result in biased estimators. Several authors have proposed methods of noise filtering, often based on recursive algorithms using prefiltered data and adaptive procedures. Contrasting with these methods, a direct method based on the asymptotic properties (for an infinite number of observations) of the experimental covariance matrix is proposed. Essentially, the procedure tries to determine the best value of the noise variance which can be used to correct the covariance matrix. This value is determined as the solution of a non-linear equation. It will be shown that this equation defines a unique solution. The advantage of this direct method against recursive filtering procedures is that no algorithmic divergence is to be feared. The paper first presents theoretical developments of the method, and then its performance is tested on several cases representing usual challenging situations associated with measurement noise: the case of two modes with close frequencies, the case of two modes with large amplitude ratios, and the case of a great number of modes.