This paper describes a continuous speech recognition system for medical diagnoses that uses a trigram model based on sequences of Japanese characters. Dictation of medical diagnoses is one of the most promising applications of speech recognition. We devised a prototype based on consonant-vowel spotting using the DP matching technique, as demonstrated at ICSLP'90, and have since improved this system by using a Japanese character trigram model , a phrase syntax and phoneme-based hidden Markov models. Speaker-dependent recognition tests have been done on 543 phrases about X-ray CT scanning. The word lexicon has about 3600 entries. The trigram model reduced the character perplexity from 11.8 to 3.6. Using the new system, 95.8% of the input phrases were correctly transcribed, compared with the 61.5% reported at ICSLP'90. These results show the effectiveness of the character trigram model for continuous speech recognition.