For speech recognition, a set of acoustic shapes (in binary format) for a given set of words, in which word-to-word differences are maximised but speaker-to-speaker differences are minimized.