Based on my observations and the details that Marco has blogged about, it seems to use a silence-detection algorithm that rapidly adjusts playback speed, so the silences are not cut out in a jarring manner, but rather played through at a typical 2.5-4x.