ED[1] writes audio frames more or less on par with the video frames but
the GOP in this file is not reordered (as it sometimes is) but in the
presentation order, so it is needed to wait almost the whole GOP until
audio frames can be correctly attached to video.
[1] The Elephants Dream
Currently it is intended rather only for simple filters that do not change
format much (especially tiling mode). When combining filters that change
video properties some issues may occur.