Its not perfect but interesting.
Check out his proof of work on reconstruction
BEHM-GAN: Bandwidth Extension of Historical Music using Generative Adversarial Networks (aalto.fi)
BEHM Gan the reconstructed with a 3k cut. Reference at bottom.
Ahhh tru! Mystery pianist AI sheeyat
DAFx-22 - Gramophone noise synthesis (aalto.fi)
send me a link to a short rec of yours. I’ll see what you sound like from grammophone era.
Oh wait… lol the project just produces noise. As it should be.
Ahh knar… it hit me. I can reverse polarity of the “denoised” version then cancel to see what it subtracted. I can then clean the noise of any true signal by reducing tonal elements and deleting transient info from it before using it on a fresh wave form to cancel out noise ultra super clean style.
I will experiment once I go ahead and set up a fake meeting to get the time.
So here the subtracted info
Holy fuck look how much musical info I save from this already superb class leading music / noise separation.
So I do final remaster touches and clean up a little by hand.
1 Like
I don’t think this method gonna make sense until I get back on the laser turntable. The transient hits on the record needles are pretty distorted which mean they get softened a bit by this method. Hmmmmm…
dayum iz dere commercial compare ?
found older project file (denoiser only without bwe high freq recuntstruction)
interestingly paramz iz set at 44100 in diz project tho I hear an artefact in da high rez
may interezt u I managed to download the noise residual directly from this project
1 Like
This by far the best available on market. Is naxos recent ho shits.
Interesting! Before it destabilize all our societal systems creating unrest which lead to the last and final war, this tech do cool stuff. I will play with this version and see what it do.
This transfer off a couple hundred dollar DJ turntable too. Fatty never get around to scratching tho.
I may already be using the project demo you link. To be sure I run another pass and they null perfect. When I got time I poss read his blurb about training your own models. I wish I had more time for this stuff.
diz my modified version with da residual noise download
https://colab.research.google.com/drive/1O7MsuUb4L4Vk5Rmaa3XO1tM3xSVNjZB-?usp=sharing
Ooh, i forgot you speak alien. 
file download at 44.1k
maybe diz actually bettah den da complicated bwe but at @22.05k
1 Like
These files probably sound wikid with fake ambient sound, touch of reverb and some tonal exciter shits. This sound like work tho…
feel like training denoiser at 44.1k and bwe up to 96k
how many SACD u have ?
Poss record your own library at 768 khz
This a bizarre but amazing device BTW.
soundz lyk success too
I find way around the temporal limits in this guy’s model. 
and create best ever transfer of this in history according to my laptop speakers.
First meeting in 4 hours. Feels like solid time investment.
Hmm in headphones a bit meh… oh well. Tinker tomorrow.
(post deleted by author)
2 Likes