Tacotron: Towards End-to-End Speech Synthesis

[jinglescode]

Paper

Link: https://arxiv.org/pdf/1703.10135.pdf
Year: 2017

Summary

• Tacotron, an end-to-end generative text-to-speech model that synthesizes speech directly from characters
• train from <text, audio> pairs, model takes characters as input and outputs raw spectrogram
• this is the first of all Tacotron development - https://google.github.io/tacotron/

Methods

• seq2seq encoder and decoder, consist of conv, attention, and GRU

Results

• outperforming in terms of naturalness
• substantially faster than sample-level autoregressive methods

Comments