Google Research is excited to announce ... TACOSPAWN! It's a TTS model that can speak in human voices that don't exist: google.github.io/tacotron/publi… We learn a prior over the speaker embedding space of a many-speaker Tacotron: arxiv.org/abs/2111.05095

Our new paper on speaker generation. We fit a probabilistic model to existing human speakers (talkers) and sample from the model to generate new ones. The goal is to generate speakers who sound like "someone you just haven't met yet" while not in fact corresponding to any person.

Want to push the frontier of machine-learned hyperrealistic speech modeling? The Tacotron team is hiring in Mountain View, CA!

Very excited to announce this paper, with my incredible co-first-author, Matt Shannon!: arxiv.org/abs/2212.03232 An attempt to revive generative (joint) modeling as a principled approach to un/semi-supervised learning in seq2seq problems.

It's really cool when a simple, pretty idea basically solves a specific problem. Very Attentive Tacotron (VAT) shows how to make transformer decoders faithfully follow their conditioning input in approximately monotonic seq2seq tasks like speech synthesis, and it's crazily robust