Deep Neural Network (DNN) as the generative model for Myanmar speech synthesis is presented in this paper. A question set for Myanmar language is proposed and used in context clustering of HMM-based speech synthesis and extracting input features for DNN-based speech synthesis. We investigated the effectiveness of precise state boundaries and coarse phone boundaries on aligning input linguistic features and output acoustic features for training DNN. The experimental results in objective evaluation show that the state boundary information give better result than phone boundary information in training DNN in terms of acoustic features,MCD, F0 and V/U, and the subjective listening tests confirm that DNN-based speech synthesis get a significant improvement over a conventional HMM-based speech synthesis in naturalness.