A method for performance evaluation of the acoustic-phonetic front end of a continuous speech recognition system, using the entropy of its output, is described. Results are given for a front end based on phonemic hidden Markov models, with various optional enhancements which have been optimised using the entropy criterion.