Welcome,
You can’t test accuracy on the Simulator, sorry. Try it with a real device and make sure that the words being spoken are present in your language model (I noticed that you check for the word “UM” but it isn’t present in your model).
Also, there will be some or possibly many null hypotheses due to noises, so there is no advantage to returning null hypotheses and/or worrying about getting null hypotheses – the important question is only whether matching speech is not detected. I would not test things out with null hypotheses being returned, since getting lots of correct negatives is likely to be confusing if what you are looking for is false negatives.