Welcome,
Sorry you’ve experiencing this issue. I’ve heard one other case of it (although it was in the opposite direction – speech was fine but files were not, weirdly) but I don’t yet know what the cause is. What happens when you raise vadThreshold to 4.4?
Another important thing is that the ‘recognitionScore’ value seems almost random.
Unfortunately, with small vocabularies the scores are extremely uninformative and even with big vocabularies they are extremely influenced by many things in the session besides what you would consider nearness of pronunciation. If you search for “score” in the forum search you’ll see many discussions about how narrow the applicability of scoring is. I wish this was better but there are actually very few cases in which scores can be used – definitely not to get some objective idea of the accuracy across multiple sessions.
If I run the WAV file, will I see the good behavior or the bad behavior? It may be a while before I have the opportunity to do so.