Deep learning-based recommendation algorithms have recently at- tracted attention due to their effectiveness at processing big data. Methods based on the variational autoencoder (VAE) are particularly promising thanks to their advantage with the data sparsity problem in recommendation tasks. However, because user traits affect the preference of recommended items, to improve the performance of VAE-based recommendation methods, it is necessary to carefully consider user traits. In this paper, we propose a method that conditions the VAE with user trait labels for switching the distributions of a generative model of latent variables. Experiments on a music recommendation task demonstrate that utilizing user trait labels estimated from tweet history leads to an improved performance and that the distribution can be changed depending on the individual traits of users.