During a loopback test, I noticed that if the previous stream ended abruptly (i.e. process terminated), then the beginning of the next stream will show a large DC offset. What's even more eyebrow-raising is that the DC offset progressively fades away, but it takes a long time for it to do so (multiple seconds) - it looks like the slow discharge of some large capacitor.

I did not yet check if it's the output side that's generating the DC offset or if it's the input side that's adding it.