Rejecto is fine at rejecting OOV utterances, but it isn’t going to work ideally in a language model is large enough that it has several other perfect rhymes for the OOV utterances if they are being spoken directly to the app, since they will compete quite evenly with Rejecto’s rejection. I would be surprised if the results were the same with 200 words, perhaps you should create a replication case to show me if you have time:
https://www.politepix.com/forums/topic/how-to-create-a-minimal-case-for-replication/
It is possible that an app concept that is based on the user saying any possible sentence and then a very large vocabulary of keywords being detected within that isn’t a good fit for offline recognition. This is basically a variation on free-form dictation, so if you can’t make the vocabulary small enough that it isn’t full of rhyming combinations for OOV words and you can’t restrict the search space of the input, it will probably work better to use a cloud-based large vocabulary dictation SDK instead of an offline speech UI SDK, and just identify and then process everything.