This study proposes an emotion clustering method based on Probabilistic Linear Discriminant Analysis (PLDA). Each emotional utterance is modeled as a GMM mean supervector. Hierarchical clustering is applied to cluster supervectors that represent similar emotions using a likelihood ratio from a PLDA model. The PLDA model can be trained with a different emotional database from the test data, with different emotion categories, speakers, or even languages. The advantage of using a PLDA model is that it identifies emotion dependent subspaces of the GMM mean supervector space. Our proposed emotion clustering based on PLDA likelihood distance improves 5-emotion clustering accuracy by 37.1% absolute compared to a baseline with Euclidean distance when PLDA model is trained with a separate set of speakers from the same database. Even when PLDA model is trained using a different database with a different language, clustering performance is improved by 11.2%.