Truth discovery is a hot research topic in the Big Data era, with the goal of identifying true values from the conflicting data provided by multiple sources on the same data items. Previously, many methods have been proposed to tackle this issue. However, none of the existing methods is a clear winner that consistently outperforms the others due to the varied characteristics of different methods. In addition, in some cases, an improved method may not even beat its original version as a result of the bias introduced by limited ground truths or different features of the applied datasets. To realize an approach that achieves better and robust overall performance, we propose to fully leverage the advantages of existing methods by extracting truth from the prediction results of these existing truth discovery methods. In particular, we first distinguish between the single-truth and multi-truth discovery problems and formally define the ensemble truth discovery problem. Then, we analyze the feasibility of the ensemble approach, and derive two models, i.e., serial model and parallel model, to implement the approach, and to further tackle the above two types of truth discovery problems. Extensive experiments over three large real-world datasets and various synthetic datasets demonstrate the effectiveness of our approach.