That is along the lines of what I was thinking, but does not make the point as cleany (they end up combining the speach-primed and non speach-primed data because participants said they percieved it as speach from the start.

They also used an ABX methodology, which forces the subject to put sounds into boxes by essentially asking if X is more like A or B, not if X is different from both.

The one I am thinking of used an odd-one-out methodology, where the subject was presented with AAX, and asked to pick which sound was different from the others (where the others were genuinly the same sound). The one I am thinking of also found a priming effect, which yours apparantly didn't.