The x-vector system maps a variable-length speech to a fixed-dimensional speaker embeddings via neural networks， and performs well in text-independent speaker verification. Here， it is applied to the text-dependent speaker verification and different x-vectors are extracted according to different contents in one sentence. In model selection， deep residual network （DRN） is used to obtain more discriminative x-vector. For a sentence with multiple words， word-dependent DRNs are trained to extract word-dependent x-vectors， which are separately fed to different backend classifiers. Finally， multiple scores are fused to obtain the final verification results. Experiments on Part Ⅲ of the RSR2015 dataset show that the proposed method can achieve equal error rate （EER） reduction of 15.34% and 19.7% for male and female， respectively.