OpenAI's Jukebox AI Produces Music in Any Style From Scratch -- Complete With Lyrics

OpenAI this week released Jukebox, a machine learning framework that generates music -- including rudimentary songs -- as raw audio in a range of genres and musical styles. From a report: Provided with a genre, artist, and lyrics as input, Jukebox outputs a new music sample produced from scratch. The code and model are available on GitHub, along with a tool to explore the generated samples. Jukebox might not be the most practical application of AI and machine learning, but as OpenAI notes, music generation pushes the boundaries of generative models. Synthesizing songs at the audio level is challenging because the sequences are quite long -- a typical 4-minute song at CD quality (44 kHz, 16-bit) has over 10 million timesteps. As a result, learning the high-level semantics of music requires models to deal with very long-range dependencies.

Re:

[Anonymous Coward]

I wish someone would furlough you from your gig of being triggered by each and every AI or Elon Musk story.

Re:

[HiThere]

Programmers often have little problem with social distancing, or even working from home. This is much less true of used car salesmen.

Some real horror shows in there.

[booboo]

E.g. https://jukebox.openai.com/?song=789015790

Re:

[DontBeAMoran]

Not sure if I should be laughing my ass off, or be afraid and running into my bunker.

Re:

[sound+vision]

The "Simon & Garfunkel" AI seems to take Sounds of Silence somewhat literally,

https://jukebox.openai.com/?so... [openai.com]

My guess is it picked up on the sharp peaks in the attack of the acoustic guitar sound, and the actual notes were too low in absolute volume for the AI to pick them up, based on whatever threshold they set.

Also, I never expected Shakira to turn experimental and go Revolution 9 halfway through a song, but here it is,

https://jukebox.openai.com/?so... [openai.com]

Re:

[sound+vision]

Straight up pop seems to do the best - Backstreet Boys and Celine Dion sound very good from the AI.

50 Cent seems to have been confused with a Jamaican,

https://jukebox.openai.com/?so... [openai.com]

Needs a lot of work.

Re:

[tonique]

"Heavy metal" "song" "Rage" made me laugh hysterically. The "lyrics" purportedly include the line "Why did Open A. I. create me?" but I can't make out the words, really.

To pretend

[edis]

For the sake of pretending. What are we doing, people, really?!!

I'm not sure what part of me is more offended.

[nightflameauto]

I'm not sure if I'm more offended as a musician, or as a human being. Those "songs" are worse than the worst bedroom recorded nonsense from the seventies and eighties before DAWs became available. Just horrible trash. I can't even think how this is an interesting use case for what passes as "AI" these days. It just gives me a huge gut-pinching WTF vibe all the way around.

Re:

[dgatwood]

It's like they somehow raised Charles Ives from the dead as an AI and turned him loose on modern music, with the rule that everything has to either be halfway atonal, or have very little resemblance to actual words, or both. My ears are bleeding after just a couple of minutes.

I'm sure it will hit #1 on the pop charts, though. It's still better than half the stuff out there these days.

Re: I'm not sure what part of me is more offended.

[Aristos Mazer]

It isnâ(TM)t great music, but it is nonetheless impressive for autogeneration, in my experience. It shows real progress over earlier attempts Iâ(TM)ve heard. Itâ(TM)s reached the âoeyour mother likes it but keep your day jobâ stage.

Re: I'm not sure what part of me is more offended

[Aristos Mazer]

Yes, but that did not generate from first principles. That was marlibs style filling in known blanks. This is actual learning and doing its own analysis. It is pretty impressive and has wider applicability than just music.

There's only one response to this..

[Rick Schumann]

..and this is it. [stickerpicker.de]

Some of these fit right in

[Taquito Sensei]

With how autotuned and overproduced popular music is these days. Some of these fit right in.

sounds (!) very interesting but...

[4wdloop]

>> On a V100, it takes about 3 hrs to fully sample 20 seconds of music

How about practicallity - I do not have a GPU as I do not play games and definitely will not have an AI-dedicated GPU like V100 ?

Well...I have heard a story about 3 Economy PhDs presenting a paper about fully parameterized model of economy. And then sombody asked about number of parameters to actually simulate "human" economy and O(n) and it turned out that there would not be enough time in the universe to simulate this model...

V100

Has a positive side

[AxisOfPleasure]

As someone who loves live rock and loves playing music myself I was a about to go off on a rant about how this is sacrilege. However, looking for a positive ( it's been another long week in lockdown! ) records companies constantly sue the arse of anyone daring to sample music without paying extortionate fees, maybe this could be the answer for those looking to make mash-ups but don't want to risk being chased down by lawyers.

Microsoft Songsmith reincarnated

[wh1pp3t]

This is about as hilarious as Microsoft’s effort to generate music to match a person’s voice.
Songsmith: http://songsmith.ms/index.html [songsmith.ms]
Result: http://www.youtube.com/watch?v... [youtube.com]

Just do Ramones songs

[mpercy]

Pretty much all the same and ~3 minutes long.

Music

[uvikal]

Interesting information. I have been trying for a long time to create background music for my projects myself, but so far all my attempts have been unsuccessful. So I found the best solution to my problem. I find the background music I need at https://taketones.com/ [taketones.com] on this resource there is always a large selection of excellent music tracks