This is a fascinating article, the first I've heard of neural nets processing anything raw, sample by sample, without a gob of translation preprocessing: https://deepmind.com/blog/wavenet-generative-model-raw-audio/
The "self generating" samples near the end are really interesting to listen to, particularly the gibberish with the lip and breath sounds, and the piano.
All of the raw processing power around lately has unleashed a torrent of pent-up creativity.
A singing neural net:
http://www.dtic.upf.edu/~mblaauw/IS2017_NPSS/
Instrument backing seem to mask much of the uncanny nature of the acapella versions. The "Soft VQ" and "Powerful VQ" female Spanish voices are particularly impressive, but I wonder if some of that impression is due to my lack of Spanish speaking? Maybe I'm not hearing odd vocalizations in the performance? Even the Spanish acapella sounds incredibly real to me.
It might not be all that long before we start seeing "no humans were harmed in the making of this album".
Paper: https://arxiv.org/pdf/1704.03809.pdf. It's based on Wavenet (see my previous post).
Not a neural net (I should have made this thread more generic) but vocal synthesis nonetheless:
https://github.com/paulbatchelor/voc_demo
Based on pink trombone (web-based vocal tract sim):
"Exciting" video demo ;-):
Babbling video demos:
The excitation source sounds really good, particularly when mixed with breathy noise.
You must be logged in to post a reply. Please log in or register for a new account.
