Many applications of computer-aided language learning system lack of evaluation & feedback of learning speech. Auto speech recognition could provide a score based on likelihood. To describe an oral English training system. Besides normal speech evaluation, error detection and prosody modification based on the result of phoneme alignment and phoneme recognition are provided by this system. Recommendations for improvement queried from correction knowledge base offer learners a timely help. Current experiment shows most mistakes made unconsciously by high-level 2nd language speaker could be corrected by it.