Seems like it could just be getting at some phonetic encoding, or even raw audio information. The grammatical and vocab transformations could be accounted for by an imperfect decoder.