Wave-Tacotron: Spectrogram-free end-to-end text-to-speech synthesis

[jinglescode]

Paper

Link: https://arxiv.org/pdf/2011.03568.pdf
Year: 2020

Summary

Contributions and Distinctions from Previous Works

• text encoding is passed to a block-autoregressive decoder using attention, producing conditioning features
• use location-sensitive attention, which was more stable than the non-content-based GMM attention
• replace ReLU activations with tanh in the pre-net
• add a skip connection over the decoder pre-net and attention layers, to give the flow direct access to the samples directly preceding the current frame