Better background noise suppression!?

When recording from a microphone, I notice that its essentially impossible to get reliable start/stop recording triggers. I have discovered that:

1) sharp sounds (e.g. foot-falls, knocks, thuds, automobile traffic) can trigger recording, even though no voices are talking.

2) Talking voices are often just a tiny bit above the ambient background, and are thus NOT enough to keep the recording going.

3) The first fraction of a second (fraction of a word) that triggered the start of recording are lost. It would be nice to capture the 10 seconds of audio that occurred, *before* the triggering event. This would require buffering; the buffer being discarded when the trigger event does not occur.

Solving 3) with buffering seems doable. I have no clue about 1 & 2, though. I guess 1) is solvable with long-time RMS averages; i.e. spikes get averaged down, however, that requires buffering to not loose the trigger even (e.g. a 3-second average would result in 3 seconds of sound being lost, if there is no buffering.

2) is really annoying....