Reply To: False positives

[Halle Winkler]

No, I don’t think there is anything wrong with your implementation. It’s expected that if your app takes in a large amount of speech and it has a very small vocabulary which spans words of a few different syllable lengths that you will encounter the OOV problem. All of the empty hypotheses demonstrate that it is doing the right thing most of the time and then occasionally substituting a vocabulary word. I think Rejecto will improve results – keep in mind that you can increase and decrease the Rejecto weighting when you experiment.