your favorite song from this ai-generated covers playlist

https://youtu.be/3Z2vU8M6CYI?t=124

Talking Heads singing lyrics from an Avril Lavigne/Nicki Minaj song:
https://jukebox.openai.com/songs/807312211

this one feels like a real song, they put the scream in the logical place and everything https://jukebox.openai.com/?song=787633465

that one has a bizarre groove in the verse

what if ariana grande came out in like 2010 https://jukebox.openai.com/?song=787635205

in the early 2000s, the man who would become the weeknd collaborated with timbaland. the audio of these sessions has never been released.... until now https://jukebox.openai.com/?song=787905712

This is the best thing that has happened this year, the bar wasn’t that high but still... pretty amazing.

that talking heads definitely impressed me, a sort of "born under punches" take on the song (which i'm not aware of because i am hugely ignorant of pop music). sounds like something byrne would actually do.

like i said upthread, i like all of the takes on "rebellious jukebox", but particularly the long one that opens up with a fucked up mangling of the bo diddley beat

the motorhead song sounds like every motorhead song. i guess that's not terribly surprising or impressive. same goes for the first 30 seconds of the agoraphobic nosebleed track.

i'm still a little unclear on what music they're training this model on. the zappa version of "stinkfoot" sounds like the "apostrophe" album in a blender (particularly, like, "excentrifugal forz" or something) but doesn't necessarily sound like his fifty billion other records. there's also a jam on roxy and elsewhere's "orange county lumber truck revisited" that i feel like could be passed off as an actual bootleg from the roxy band. "tom fowler's" bass on it is particularly impressive. i also really like the hall & oates song - well, you feed a hall and oates song in not surprising that the results are good!

i want to hear more of bill withers' version of "let it be"! the first 40 seconds of the new caruso recording is good too, the rest a bit of a mess. i want to hear that duet between caruso and elvis mad magazine promised me back in the '50s!

pplains i get what you're saying but in another sense this is sort of the appropriate soundtrack for the world we live in?

Absolutely. Another point against it!

That Elvis one of Come On Eileen.... The way the first part was incoherent speech patter followed by uplifting vocals by The Stamps.... I wasn't prepared to learn How Deep My Uncanny Valley was last night.

times like this I wish I was better at audacity because on the linkin park song, if you just added a second instance of chorus after the second verse, and repeated the opening guitar riff somewhere (outro and/or after the chorus(es), you could pass it off as an actual, non-AI-generated song

also I don't know what the weeknd songs were trained on (or why this thinks he was a one-hit wonder with "high for this") but whatever it is makes some plausible beats https://jukebox.openai.com/?song=787902766

a lot of the lyrics they've done examples with are pretty weird choices, probably just chosen fairly randomly from their huge dataset (which was 1.2 million songs, half of those in english) so i'd expect there were a substantial enough number of songs for each of the artists they have 'styles' learned for

another weird thing is a huge number of the unseen lyrics examples are reggae lyrics, which seems to give very weird results from all the patois in the lyrics

http://jukebox.openai.com/?song=787982923

The only Lou Reed song is big and dumb in a perfect Lou Reed way.

i was curious as to what they would do with can but it turns out they do have an attempt at "turtles have short legs" up there and it's awful

the yoko ono is much better, some great record scratching on that one

https://jukebox.openai.com/?song=807320566

so they tried to do a bunch of different versions of "beat on the brat", none of them sound a goddamn thing like the ramones and most of them are fucking awful, i have to wonder what they trained it on. i guess the takeaway here is that most punk music doesn't sound anything like the ramones and is usually much worse.

here are the versions i found most interesting.

https://jukebox.openai.com/?song=787896025
https://jukebox.openai.com/?song=787909846
https://jukebox.openai.com/?song=787911817
https://jukebox.openai.com/?song=787916911
https://jukebox.openai.com/?song=787913575

you would think the ramones would be easier to emulate

yeah, surprised the hell out of me too. hate to go for the nerd chestnuts but it seems like xkcd 1425 applies here.

(yes, i had to google it, and while it's nice that they actually put a woman in their strip as a programmer, i admit to being a bit deflated to learn that the alt tags identify her as "Woman Programmer")

can someone point me to a description of how exactly this technology works and what it's doing, and also how it relates to the Vocal Synthesis thread which is obv using different tech and has clips that sound somewhat more realistic?

anyway, this is supposed to be blues in the style of Chris Rea - it isn't, but it's pretty great anyway

https://jukebox.openai.com/?song=787528894

the paper: https://cdn.openai.com/papers/jukebox.pdf

lee - what i think that ai does is just generates something that sounds like any given person's voice, speaking or singing, acapella. the only one of those videos that has a musical background is the sinatra one, and i'm not sure but it sounds like for that they generated an acapella vocal track and inserted it on top of a pre-existing background. what i don't know is if the pitches in the sinatra version were pre-defined. for all of the others, though, pitch isn't a consideration.

anyway, in jukebox, none of the pitches are predefined, the software doesn't perfectly know how to differentiate between "vocals" and all the different instruments, between speech and applause, and rather than strictly generating sounds in a particularly "voice" that match a predefined text - this is part of what it does - it also attempts to create novel musical content in a similar style to the work it's emulating. so the eight different versions of "industrial estate", none of them actually contain the melody (such as it is) of "industrial estate", but all of them sound like songs by the fall, they have some of the, uh, accidents of early Fall songs, recognizable drums and guitars and vocals and so forth.

ok now I'm freaked the fuck out because I am pretty sure I am the only person alive who remembers the source track to this

https://jukebox.openai.com/?song=787991059

Had no idea that AC/DC was a sludgy stoner band now

https://jukebox.openai.com/songs/802878241

A demo from Johnny Cash's brief early-'80s flirtation with no wave.

https://jukebox.openai.com/songs/799894093

Almost every Muppets track on there is an absolutely nightmarish audio collage designed to melt your brain

https://jukebox.openai.com/songs/787623133

Thx for the explanations

I think the most amazing thing I learned from all of this is the fact that The Verve Pipe actually put out a Children's album in 2009:
https://en.wikipedia.org/wiki/A_Family_Album

And followed it up with a second one in 2013.

it talks about why it's bad to pass women back and forth, lest you be guilt stricken, sobbing.

can't wait for Leyland Kirby to release 200 LPs of this shit

Okay, I just died

https://jukebox.openai.com/songs/787717459

Damnit, I can't believe I dropped my 'I Just Called to Say I Love You' cassingle in a puddle

https://jukebox.openai.com/songs/787731991

what the FUCK at muppets Baby Shark

since i don't have kids or hang out with people who do, this is the only version of "baby shark" i have ever heard and i am going to believe for the rest of my life that "baby shark" sounds exactly like that

You're in an enviable position. I would recommend that you remain there.

lol, p sure the robot is smoking crack with some of these. Here's 'Nirvana'.

https://jukebox.openai.com/songs/799441342

Also p sure listening to these things is actively eroding the speech center of my brain.