This paper proposes a technique for aggregating mixtures of probabilistic principal component analyzers, which are a powerful probabilistic generative model for coping with a high-dimensional, non linear, data set. Aggregation is carried out through Bayesian estimation with a specific prior and an original variational scheme. We demonstrate how such models may be aggregated by accessing model parameters only, rather than original data, which can be advantageous for learning from distributed data sets. Experimental results illustrate the effectiveness of the proposal.