We introduce a deep latent recommender system (deepLTRS) for imputing missing ratings based on the observed ratings and product reviews. Our approach extends a standard variational autoen-coder architecture associated with deep latent variable models in order to account for both the ordinal entries and the text entered by users to score and review products. DeepLTRS assumes a latent representation of both users and products, allowing a natural visualisation of the positioning of users in relation to products. Numerical experiments on simulated and real-world data sets demonstrate that DeepLTRS outperforms the state-of-the-art, in particular in contexts of extreme data sparsity.