fatty 
fatty 
itz not full band (workz at 22k zample rate) but think it ztill beat izorape
full track
1 Like
They sound good but the hiss annoys da sheeyat outta moi. Screenshot of da page pozz as it wantz login? 
I’m curious as to wut they say.
Oh I see. 2nd track is AI noise reduced or sumsheeyat?
2 Likes
Yeah, the second version of track one is actually cleaner than Izotope after a quick denoise and decrackle. (Decrackle is not at full strength when I did it but I don’t think that would help.)
tru firzt trackz are fatty teflon input
fund diz mofo on dizcord n quite incredible dat he rezearchz exactly fatty uze caze
or iz diz da fatty ?
3 Likes
I’d like to make it work at 44.1 khz
1 Like
Daim da early Ho sounds smoother with hiss
1 Like
HOLY FUCK! Finally somebody doing research of value!
I will explore while nodding in my next set of meetings. One with a dude who wrote some of Trum*'s speeches so I can go ahead an zone out there.
1 Like
If I were ever to be educated it would be at this university for sure.
1 Like
I dun read the paper but have a 96khz transfer I’ll try of the valse
It look like it maybe give me 44100
When file upload, I’ll see if I can manually shift the sample rate up or down. Also you should copy their code before they package this into some unusable android app to raise money off of they never actually plan to monetize.
1 Like
This is extraordinary!!
Expecting AI tools to make this happen is why I didn’t want to waste time learning noise reduction software and just concentrated on getting nice transfers. It seem to be way to go.
The problems introduced by their method super easy to mitigate too. I add background noise so it dun sound so dead, I take their 22khz output and make a left and right channel out of them with some mild cross feed and tweak phase a bit. 95% of weirdness gone in a minute or two.
Damn!~
Next I’ll work on fixing impulse drop and maybe adding reverbs to better fill dead space but I dun have those plugins on this machine.
why dey trained at ghey 22050
“better NEVER change”
I fee lyk changin 
1 Like
Manually changing the sample rate to 96khz cauzed an output at 48k but at wrong speed.
#codemystery
they train convolution filterz on time domain not frequency
filterz only zaw music @22050 during training
if u give him 44100 muzic, model will juz think diz normal muzic at 
zpeed
1 Like
da fatty zhould have ztayed at facerape so we cud use the supercluster
1 Like
Damn… is probably ok for my sheeyats tho, ~99% of everything about 5 or 10k noise anyway.
Ooh I have so many ideas! Thanks K-nar!!
1 Like
Bro, it’s all in the hiss
I like the noise reduced version better just like Ichika Matsumoto looks better in skintight leather than naked
I got some ideas on how to maximize this but yeah a little unnerving when listening to raw outputs. Hold da door. Lemme try the hung rapper 13 on this.
Gonna use the ultra fucked one with the super high treble push.
Hmm not enough music signal in there. This basically give it some good result and some that sound like demon possessed insects whispering spells at me.
I think the process to make audiophile happy wud be
make a zecret hairless denoised version
add good tazte warm hizz
zell at 96hz 24 bit
1 Like
Randomly a lot of “remasters” are just people using 44.1k stuff transferred in the 80s brought down to nubs with some fake pleasant sounding noise used to fill the background. Half the horowitz remasters from an era when no masters exist.
They should have splurged on bit depth tho. If the k-nar want to become a thousandaire he should recreate their project using 24bit sources.
1 Like