What is being transmitted? For voice, the selector, or 'voter' calculates the 'SNR' based on the power present in, say, 2 frequency bands (of the demodulated signal): the low frequency band (0..2700Hz, say) and the high frequency band (all signal above xx Hz is assumed to be noise).
Real voters use a number of parameters derived from the signal to choose the best signal, e.g. articulation index/depth, signal autocorrellation, SNR', etc.
For SNR of non-speech signals, e.g. data, you might define some method of measuring what you think the signal should be (e.g. frequency at the bitrate and some (sub-) harmonics), and calculate the power. Likewise for the expected noise spectrum part of the signal (everything that's not in the 'signal' part of the spectrum).