The duration of each phoneme is extracted for several emotions. Information on phonemes and their duration are used to train a Variational AutoEncoder (VAE) to create a latent space z which represents emotion information. The loss functions that were used for that purpose are reconstruction loss, Kullback-Leibler (KL) divergence and multiclass N pair loss. Test samples are classified using the nearest neighbor criterion between their representation and the clusters associated to each emotion, as estimated from training data. To evaluate the models two metrics were used: emotion recognition accuracy and the consistency of the clusters of the latent space.