Learning interpretable control dimensions for speech synthesis by using external data

The following speech samples relate to a paper submitted for review at INTERSPEECH 2018

Control of emotion

Here we set the control vector to one-hot vectors for each emotion

Paragraphs

We present 3 systems, for pairwise paragraph comparison

• DNN-C: controllable SPSS voice, controlled by one of the authors using a simple UI
• DNN-R: controllable SPPS voice, with random control values
• DNN-B: baseline SPSS voice