A Background

In multi-scale RNNs (Chung et al., 2016), faster features predict when to hand control to the slower features. CHiVE (Kenter et al., 2019) is a temporally abstract hierarchy for speech synthesis.