Unfortunately this doesn't really work, as the model is not limited in it's decoding vocabulary.

Does anyone have other suggestions that may work in this space?