In this paper we address the recovery of a finite stream of Dirac pulses from noisy lowpass-filtered samples in the discrete-time setting. While this problem has been successfully addressed for the noise-free case using the concept of signals with finite rate of innovation, such techniques are not efficient in the presence of noise. In the FRI framework, the determination of the location of Dirac pulses is based on the singular value decomposition of a matrix whose rank in the noise-free case equals the number of Dirac pulses and the signal can be related to the non zero singular values. However, in noisy situations this matrix becomes full rank and the singular value decomposition is subject to subspace swap, meaning some singular values associated with noise become larger than some values related to the signal. This phenomenon has been recognized as the reason for performance breakdown in the method. The goal of this paper is to propose a novel algorithm that limits the alteration of these singular values in the presence of noise, thus significantly improving the estimation of Dirac pulses.