We consider the speech enhancement problem in a moving car through a blind source separation scheme involving two closely spaced microphones. We propose the use of a double fast Newton transversal filter algorithm (DFNTF) to estimate and suppress coherent noise components from speech, and a model of signal mixtures able to represent correctly the effect of microphones spacing. We also consider the realistic case where the noises at the sensor inputs contain non-coherent components. The simulation results show that the DFNTF algorithm, when controlled by a Voice Activity Detector (VAD), is able to fully cancel the correlated noise components from speech.