After more and more genome sequencing projects, like the “Human Genome Project”, the prediction of genes, including their coding region and their regulatory region, has received a lot of attention. Softwares such as GENSCAN and GeneMark are powerful, but still do not meet the requirement of the practical application. The GENSCAN predicts exons accurately, if the sequences predicted does not have insertions and deletions in their coding regions. But if it does have, even only one, the prediction could be disturbed seriously and satisfactory results can not be obtained. A hidden Markov model with states of deletions, insertions and main state is introduced to find the error of deletions and insertions. The result shows that sensitivity and specificity in exon level are both higher than 84% on the Burset/Guigò test data set.