Simple point that is widely ignored: you need hard real-time capability to offer any meaningful “soft” real-time. Let’s suppose you say that you can drop up to N frames a minute or K< N frames over any 10 second period from a nominal H>N frames per minute feed. This is a typical soft real-time task – although stated here with greater precision than usual. Now what happens at time T when we notice that K-1 frames have been dropped over the last 9 seconds? Now we have a requirement to drop no more than 1 frame in the next second.