Abstract: In order to improve the accuracy rating,a pronunciation evaluation method based on the phoneme model perception was proposed.The phoneme model perception of variation pronunciation was defined by the expectation of the difference between the log-posterior probability which is computed by the different corpus sets,and it was the basis of generating the corresponding speech recognition candidate model set.The results of experiments shows that the proposed method reduced the size of speech recognition model set about 95%;In the non-native speech database,the proposed method could achieve a correlation of 0.828 with the expert scoring,with the errors detection rate of 70.8% for Chinese initials and vowels of 42.5%for tone,which were both better than other methods.