My colleague has the idea that it might be an out-of-sync issue where the gap between audio and video data gets too high so that it still shows the buffered video-data whereas the audio data is missing and then starts a new publishing of the resnyced stream. Could this be the case or is that unlikely? What can we do to prevent this behavior?
If there are two recordings, that indicates the connection was broken. You should see that in logs on both sides. FMLE will start trying to reconnect right away. That indicates a network issue, and your confidence in the encoder's connection doesn't seem very high. Wowza will not break recordings in this configuration unless the stream is stopped and started, either on purpose or because of network problems.
One reason in this case, using Flash as the live encoder, might be the audio is not loud enough for the default Microphone silence level. Take a look
at this article.
Otherwise, network issues, as seems indicated here, can be the cause. Audio and video are streamed separately, so there are many scenarios where audio and video are out of sync, or in your case missing for some period as the result of network problems upstream from Wowza, which just reflects and records the stream it gets.