This recipe is the speaker diarization recipe for The First DIHARD Speech Diarization Challenge (DIHARD 2018). There are two tracks in the DIHARD 2018 competition , one uses oracle SAD (track1) and the other required that SAD was performed from scratch (track2). This script is for track1. The recipe is closely based on the following paper: http://www.danielpovey.com/files/2018_interspeech_dihard.pdf but doesn't contain the VB refinement. The whole system mainly contains full-covariance GMM-UBM, i-vector extractor (T-matrix), PLDA scoring and agglomerative hierarchical clustering. The VoxCeleb datasets are used for training i-vectors and PLDA. The development set of the DIHARD 2018 competition is used as validation set to tune parameters. The system is tested on the DIHARD 2018 evaluation set.