Phonological structures for speech recognition