The easiest approach would just be averaging – since noise on these three microphones should at least partially be uncorrelated, that would improve SNR.
However, that requires that the three microphones are
- synchronously sampled (i.e. the $n4th sample from each microphone happened "exactly" at the same time) and
- in the same distance from the desired source of sound.
Since both 1. and 2. are usually not fulfilled, you'll have to look into adapting to that – the result is a beamformer. This might be a term often linked to regularly spaced microphone arrays, but that's not a mathematical requirement – you'd just start by detecting the phase/time offset between the strongest (or, better: the desired) source of sound on each of your microphone channels, and then align (read: fractional delay) the channels in a manner that maximizes SNR for that. This inherently leads to a geometric beam shape, though it usually is but a singular value decomposition of a receive signal matrix (that's where "beamforming" and "MIMO processing" meet).