This paper introduces novel two-channel a priori Signal-to-Noise Ratio (SNR) estimators for use in frequency-domain speech enhancement algorithms. The SNR estimation is based on statistics of the noisy phase difference between two microphones in each frequency bin. Namely, the corresponding probability distribution is derived assuming a complex Gaussian model, and is written in terms of the SNR only. This shifts the two-channel SNR estimation problem into a classical statistical parameter estimation problem, which we propose to solve via an online version of the Method of Moments. The usefulness of the resulting approach is demonstrated in a realistic speech enhancement experiment, where a simple Wiener-based solution is objectively found to perform competitively when compared with other recent two-channel noise reduction techniques.