This paper describes a method for assessing the likely accuracy of a speaker independent isolated word recogniser on a new vocabulary, without the use of field trials or database collections. An algorithm is introduced which attempts to predict vocabulary difficulty using a word confusability measure. This is based on the likely pronunciation of words, coupled with information about likely phone confusions. The power of this technique to predict recogniser accuracy is tested through comparison with the measured performance of an actual recogniser on the 'Macrophone' telephony speech corpus.