In this paper, we use extended letter-to-sound rules for automatic mispronunciation detection, aiming at checking pronunciation errors made by Chinese learners of English. The knowledge-based approach is used to generate extended pronunciation lexicon and incorporated into the HMM-based mispronunciation detection system. The pronunciation errors lead to misunderstanding of a word are expected to be identified. The TIMIT text prompts are used to collect data from Chinese university students, and the test set includes a total of 1900 sentences. Experiments show that the F-measure is about 0.86 at word level and about 0.91 at phone level. The system shows a high degree of accuracy in classifying correct and erroneous pronunciation.