The focus of this paper is the estimation of the dynamic transfer function between two outputs of a linear system subjected to an uncontrolled and generally unknown excitation, and accounting for possible uncorrelated noise present at both outputs. Several applications of this case exist in the passive identification of dynamic systems including the health monitoring and/or non-destructive evaluation of structures subjected to natural "ambient" excitations. It is well known that noise-robust transfer function estimation of a single-input-single-output system can be achieved by a normalized cross-power spectrum operation. This paper shows that, for the subject case of a dual-output system, particular caution must be placed in the choice of the normalization factor to apply to the cross-power spectrum of the two outputs. In particular, an "inter-segment" averaging method is proposed for the normalization factor in combination with the classical "intra-segment" averaging of the cross-power spectrum in order to estimate the transfer function between the two outputs without the influence of the excitation spectrum and of the uncorrelated noise at the two receivers. Validating results are presented for synthetic signals and for experimental signals from an application to high-speed ultrasonic rail inspection exploiting the train wheels as the "ambient" excitation.