The paper describes a method for recognizing large vocabulary isolated words in two steps: a preclassification step in which cohorts are scored, and a fine matching step in which the words belonging to a number of best scoring cohorts are compared with the input. The two steps use essentially the same algorithm with a different representation of acoustic knowledges cohorts are represented in terms of gross phonetic classes, while words in terms of diphones. Distance measures from gross classes are simply obtained by taking the minimum of distance measures from the diphones included in the gross classes. Recognition tests on words belonging to a 10,000 word vocabulary indicate that a reduction of computational load to slightly over 20% of that needed to compare the entire vocabulary is obtained, with a very small loss in recognition accuracy.