The error rate evaluation is very important in building an Automatic Speech Recognition (ASR) system. The conventional algorithm for Word Error Rate (WER) evaluation is based on the minimum error rate. The improved WER algorithm is proposed on the basis of minimum error rate time information, which makes the WER evaluation on the overall system more accurate. Then the acoustic model then can be improved according to the evaluation.The applications of the new algorithm in building a practical ASR system is also presented.