Erratum
-------

The paper "Principles for learning controllable TTS from annotated and latent variation" published at Interspeech 2017 contains the following error:

* Page 3958 of the paper in the proceedings gives the learning rate of the latent vectors as $2\cdot10^{-3}$. This should be $2\cdot10^{-4}$. The error has been corrected in the first author's online postprint of the article.