First, message resources are loosely coupled to audio resources -- you take the module/noun/verb/case/sequence numbers, calculate a hash number for it, and if there's an audio resource with the same number, that's what gets played. And it's exactly the same with lip syncing for that audio.
Therefore, if you only edit a single value in a message resource, there's no real reason for the audio to completely break beyond SCI Companion maybe being overly eager to help, perhaps because by changing a verb value you've changed that line's hash and SCI Companion feels a need to adjust the audio samples to match. And it might be messing up at that point. If it didn't try to do these things, the worst you could expect is that you'd hear an entirely different line, or none at all.
But all this is me theorizing really hard before I go to bed so take it with a grain of salt.
Just to be sure, you might want to check if you're using the latest version of SCI Companion. The one on scicompanion.com is somewhat outdated. If nothing else, the latest version (available from the link below in my signature) has reference support so you don't get mysterious blank entries where a line has multiple different nouns/verbs/etc.