Amazing new methods in speech / singing synthesis

Discuss music production with Ableton Live.
Post Reply
Angstrom
Posts: 14923
Joined: Mon Oct 04, 2004 2:22 pm
Contact:

Amazing new methods in speech / singing synthesis

Post by Angstrom » Sun Apr 30, 2017 9:51 am

It's not a product yet but I suspect one of the big players will snap up a license to this code.

Demos.
http://www.dtic.upf.edu/~mblaauw/IS2017_NPSS/

Techno blurb.
http://www.creativeai.net/posts/W2C3baX ... ynthesizer

[jur]
Site Admin
Posts: 5405
Joined: Mon Jun 01, 2015 3:04 pm
Location: Ableton

Re: Amazing new methods in speech / singing synthesis

Post by [jur] » Sun Apr 30, 2017 2:32 pm

8O that's really impressive!
Ableton Forum Moderator

TabSel
Posts: 90
Joined: Thu Jan 05, 2012 7:05 am

Re: Amazing new methods in speech / singing synthesis

Post by TabSel » Sun Apr 30, 2017 4:14 pm

Absolutely impressive. VST NOW! ;)

Martin Gifford
Posts: 439
Joined: Mon Jun 14, 2010 12:48 am

Re: Amazing new methods in speech / singing synthesis

Post by Martin Gifford » Mon May 01, 2017 11:45 am

Is it real? If so, then they'll get bought out by some kind of robotics multinational. If it's an entirely new approach, then they are geniuses.

Angstrom
Posts: 14923
Joined: Mon Oct 04, 2004 2:22 pm
Contact:

Re: Amazing new methods in speech / singing synthesis

Post by Angstrom » Mon May 01, 2017 4:20 pm

Martin Gifford wrote:Is it real? If so, then they'll get bought out by some kind of robotics multinational. If it's an entirely new approach, then they are geniuses.
its real. It's a neural network. If you have seen some of the Deep Dream images which look like psychedelic pattern matched slug based artworks then this is related. It's baby AI computation. It takes a curated set of input data, and uses that to create probable outputs. In deep dream they trained the net on image libraries of slugs, dogs, architecture, then it was interpreting every input as slugdogs. But that was them messing around. The real purpose is to create good simulations which copy the characteristics of the bulk input. Adobe licensed some of this tech to create a voice editor demoed last year.
This code uses data sets of phonemes, and uses a neural net to make output through a convolving vocoder.

I think this singing synth is likely to be licensed by all kinds of people, someone like iZotope can package it up as a VST. Others will use it in other ways. I very much doubt that robotics is the primary profitable licensing market right now.

Some more on wavenet tech https://deepmind.com/blog/wavenet-gener ... raw-audio/

TomKern
Posts: 358
Joined: Mon Dec 05, 2016 7:08 pm

Re: Amazing new methods in speech / singing synthesis

Post by TomKern » Tue May 02, 2017 8:13 am

Angstrom wrote: I very much doubt that robotics is the primary profitable licensing market right now.
Maybe not right now, but it could be the main one in just a few years. And what about Siri, Cortana etc.?! Or NPCs in games? Could be huge!

Post Reply