Continuous speech recognition with automatically selected acoustic prototypes obtained by either bootstrapping or clustering.