Text-to-speech systems recently achieved almost indistinguishable quality\nfrom human speech. However, the prosody of those systems is generally flatter\nthan natural speech, producing samples with low expressiveness. Disentanglement\nof speaker id and prosody is crucial in text-to-speech systems to improve on\nnaturalness and produce more variable syntheses. This paper proposes a new\nneural text-to-speech model that approaches the disentanglement problem by\nconditioning a Tacotron2-like architecture on flow-normalized speaker\nembeddings, and by substituting the reference encoder with a new learned latent\ndistribution responsible for modeling the intra-sentence variability due to the\nprosody. By removing the reference encoder dependency, the speaker-leakage\nproblem typically happening in this kind of systems disappears, producing more\ndistinctive syntheses at inference time. The new model achieves significantly\nhigher prosody variance than the baseline in a set of quantitative prosody\nfeatures, as well as higher speaker distinctiveness, without decreasing the\nspeaker intelligibility. Finally, we observe that the normalized speaker\nembeddings enable much richer speaker interpolations, substantially improving\nthe distinctiveness of the new interpolated speakers.\n