Handling live translation of cues to WebVTT

I previously sent this email to the TimedText WG,<http://lists.w3.org/Archives/Public/public-tt/2014Jan/0005.html> but it was suggested by Silvia that I should bring the conversation here.

Two suggestions were made on that mailing list:

It may be better to set the end time to a smaller value, so if something unexpected happens, the cue will still go away eventually. We could probably handle this by setting the end time to 10 or 20 seconds, and extending it if we don't find the correct end time before the cue is supposed to end, like so:
1
00:00:05 --> 00:00:20
First cue

NOTE We don't know when the cue ends yet at the 20 second mark, so we add another 20 seconds
1
00:00:05 --> 00:00:40
First cue

NOTE Now we know the correct end time
1
00:00:05 --> 00:00:25
First cue
The other suggestion was just replacing the entire document to update a cue. This would be useful because the document wouldn't increase in size forever. The downside (my opinion) is that it would make the client more complicated and less efficient (because we'd need to do polling). I don't think unbounded file size would actually be a problem in WebVTT though, since the UA isn't required to hold onto the original document. WebKit and Chromium/Blink both don't -- They parse the file contents as they receive it, then throw it away.