为了提高发音质量判别精度,提出了一种基于音素模型感知度的发音质量评价方法。它采用不同语音样本集合下样本声学特征的对数后验概率期望差作为音素模型对变异发音的感知度,并以此为基础,生成各音素对应的识别模型候选集。实验表明,所提出的方法使语音识别网络候选音素模型集合尺寸减少约95%;在非母语语音数据库上,该方法评分与人工专家打分相关性为0.828,基于该方法得到的声韵母错误检出率为70.8%,声调错误检出率为42.5%,均优于其它方法。 In order to improve the discrimination accuracy of pronunciation quality, a pronunciation quality evaluation method based on phoneme model perception is proposed. It uses the expectation difference of the logarithm posterior probability of the sample acoustic features under different speech sample sets as the perception of the variant pronunciation of the phoneme model and generates the candidate set of recognition models corresponding to each phoneme based on the expectation difference. Experiments show that the proposed method reduces the size of the set of speech recognition network candidate phoneme models by about 95%. On the non-native speech database, the correlation between this method score and artificial expert score is 0.828. Based on this method, the vowel consonant error detection The rate was 70.8%, and the tone error detection rate was 42.5%, which were all better than other methods.