English and Japanese are quite different languages both phonetically and linguistically and it is often very difficult for Japanese students to master English pronunciation. To help students improve their pronunciation proficiency, a Japanese national project of "Advanced Utilization of Multimedia for Education" has started in 2000 and under this project, a large database of English words and sentences read by 200 Japanese students was built mainly for CALL system development. This paper describes a corpus-based analysis and comparison of American English (AE) and Japanese English (JE) by using WSJ database and the new JE database. Here, Hidden Markov Models (HMMs), which are widely-used acoustic modeling techniques of speech recognition, were firstly made for individual phonemes in the two kinds of English, and then, a tree diagram was drawn for the entire phonemes of each HMM set. The analysis and comparison of the two trees showed many interesting characteristics of JE, some of which are well-known habits observed in JE pronunciation. The authors consider that this study showed statistical differences between AE and JE in view of the entire phonemic system of English for the first time.