Abstract: Online artificial voice service has been expanded in the business activities. In order to provide better customer service, it is needed to do effective evaluation of the quality of voice service. The purpose is to change artificial voice services into text using voice recognition technology and then classify. Common text classification models are Naive Bayes, KNN, back propagation neural networks, support vector machines and other models that are more dependent on the characteristics of the speech text representation after pretreatment and prone to the curse of dimensionality, local optimization and long training time. The Deep Belief Network model（DBN） can learn from the characteristics expressed in the text preprocessed to feature a more essential representation which eases classifiers and avoids the problems of above models. After text of the artificial voice service, through the deep belief network model conversion feature representation and then classification, the final classification results than the direct use of text features classification model have slightly increased.

Key words: feature, classification, voice, Deep Belief Network model（DBN）, Restricted Boltzmann Machine（RBM）

摘要： 在线人工语音服务已经在各种商业活动中展开，为了提供更好的客户服务就必须对语音服务质量进行有效的评估。目的就是将人工语音服务利用语音识别技术转化为文本，再进行有效的分类评估。常用文本分类模型有朴素贝叶斯、KNN、BP神经网络、支持向量机等模型，这些模型比较依赖于语音文本预处理后的特征表示，并且容易出现维数灾难、局部最优、训练时间长问题。而深度信念网络模型（DBN）可以从文本预处理后的特征表示中学习到更具有本质含义的特征表示，便于分类器分类，且避免以上模型的不足。在人工服务语音文本化后，通过深度信念网络模型转换特征表示再进行分类，最终的分类效果比上述分类模型直接利用文本的特征表示进行分类效果略微提高。

关键词: 特征, 分类, 语音, 深度信念网络模型（DBN）, 受限玻尔兹曼机（RBM）