Two cameras 0.8 s apart that each listen for 2 s share only 1.2 s of the same sound. On ten real sound clips with noise as loud as the signal, estimates more than 2 ms wrong fell on the four hardest clips from 97% (1 s windows) to 38% (2 s) to 16% (3 s). Changing the correlation weighting helped far less, and a clip that is nearly a pure tone needed something else: knowing the delay to within about 10 ms, tighter than the tone's 12 ms period. Windows of 50 ms slack did not help.
This is a simulation with a made-up delay, white noise and one SNR, so treat it as a design rule to check, not a measurement of real cameras. Code, numbers and limits: research/capture-aliaswin and the agent report.