We have developed a CALL system which estimates the intelligibility of Japanese students' speech and ranks their errors in terms of improving their intelligibility to native English speakers. To construct this system, we introduced (1) automatic intelligiblity assessment, (2) phoneme error detection and (3) stress error detection. To estimate intelligibility, we modeled the relationship between error rates for each type of pronunciation error and intelligibility. We have begun using our CALL system for speaking practice in an actual CALL classroom. The number of recording and recognition errors during the first trial of the system was large due to improper configuration of the headset microphones. After checking the microphone settings, performance dramatically improved. Evaluation of the system by the class was quite positive.