Postpartum depression (PPD) is a depressive disorder with peripartum onset, which brings heavy burden to individuals and their families. In this paper, we propose to detect PPD in depressed people via voices. We used openSMILE for feature extraction, selected Sequential Floating Forward Selection (SFFS) algorithm for feature selection, tried different settings of features, set 5-fold cross validation and applied Support Vector Machine (SVM) on Weka for training and testing different models. The best predictive performance among our models is 69%, which suggests that the speech features could be used as a potential behavioral indicator for identifying PPD in depression. We also found that a combined impact of features and content of questions contribute to the prediction. After dimension reduction, the average value of F-measure was increased 5.2%, and the precision of PPD was rose to 75%. Comparing with demographic questions, the features of emotional induction questions have better predictive effects.