A.I. "Jazz"

[AROIOS]

The short piece at 8:09 is actually enjoyable, more so than 99% of the garbage on most charts today.

If that's any indication of our AI overlord's potential, I'm all for it.
 

 

 

[JoJoB3]

33 minutes ago, AROIOS said:

The short piece at 8:09 is actually enjoyable, more so than 99% of the garbage on most charts today.

If that's any indication of our AI overlord's potential, I'm all for it.
 

 

 

"...more so than 99% of the garbage on most charts today" 
 

This?  I'd have to question if you understand Bill Evans or jazz for that matter. The examples here are garbage that only serves to go along with this A.D.D. fueled video (complete with unfunny jokes).  Watching this was a waste of time.  I'm sorry.

[Baldwin Funster]

It's not that remarkable that AI can "compose" some kind of music.

But if you come home and your robot is LISTENING to music you better watch out. Next up is robot REVOLUTION. 

 

My robot made me type this....

[BenWaB3]

Sorry, sounds like hammered dogshit. I'm also sorry I spent the whole 13+ minutes waiting for it to miraculously start sounding good.

[elsongs]

It's definitely sound, but it's not music.

[Docbop]

Sound like piano bar in Twin Sh17s Iowa Holiday Inn. 

[tapes]

Anyone else already sick of the whole "A.I. does this & that" (and usually not very succesfully)? Like, I'm sorry - there will never be another Bill Evans, or Miles Davis, and just vaguely replicating some sonic character isn't it.

[stoken6]

4 hours ago, elsongs said:

It's definitely sound, but it's not music.

This. It sounds like it was (very intensively) trained on "sound" not "music". (You can hear that it decides to add the decay-tail of one piano chord a few milliseconds after the initial portion of another - because nothing is telling the AI "this is a note being played NOW")

 

There is a field of natural language research that focuses on "attention" (which words in a prompt relate to other words). This is used in GPT solutions as a pre-processor before the generative part. Something similar is lacking in music AI (at least all the music AI I've heard) - there's no concept of short/medium/long-term progression: call/response phrases,  tension/release, verse/chorus (or movement, whatever) structure, etc. And without it, AI music just sounds meandering and aimless.

 

Cheers, Mike.

 

 

[Stefan011]

3 minutes ago, stoken6 said:

This. It sounds like it was (very intensively) trained on "sound" not "music". (You can hear that it decides to add the decay-tail of one piano chord a few milliseconds after the initial portion of another - because nothing is telling the AI "this is a note being played NOW")

 

There is a field of natural language research that focuses on "attention" (which words in a prompt relate to other words). This is used in GPT solutions as a pre-processor before the generative part. Something similar is lacking in music AI (at least all the music AI I've heard) - there's no concept of short/medium/long-term progression: call/response phrases,  tension/release, verse/chorus (or movement, whatever) structure, etc. And without it, AI music just sounds meandering and aimless.

 

Cheers, Mike.

 

 

This. The machine learning using notation and midi would bring us much closer. A lot of jazz theory is basically puzzles to be solved in a certain way. AI has already been trained to identify writers (Shakespeare for example). It can easily analyze the way Bill Evans approached certain chord progressions. Next thing is microtiming and timbral quality but again nothing impossible.

 

Spotify algorithm already recommends some kind of Swedish pseudo jazz easy listening to most casual listeners. AI can get there in a few years.

[tapes]

2 hours ago, Stefan011 said:

Spotify algorithm already recommends some kind of Swedish pseudo jazz easy listening to most casual listeners. AI can get there in a few years.

 

A couple of the most popular search terms on YouTube now seem to be "jazz music" or "jazz for studying". While this is very interesting, I'd say that for a lot of casual listeners music is mostly sound, rhythm (the beat) and lyrics. But they definitely don't have a clue what's going on under the hood. So in that sense something that "sounds jazzy" is jazz in their mind. Even though the core aspect of it - improvisation - might be missing completely. A lot of people think that Smooth Jazz is jazz, even though it is technically instrumental R&B.

 

(By the way, don't knock them Swedes - a lot of great jazz came from there, and the Nordics/Scandinavia 😄)

[Thethirdapple]

Quote

Computational expressive music performance studies the analysis and characterisation of the deviations that a musician introduces when performing a musical piece. It has been studied in a classical context where timing and dynamic deviations are modeled using machine learning techniques. In jazz music, work has been done previously on the study of ornament prediction in guitar performance, as well as in saxophone expressive modeling. However, little work has been done on expressive ensemble perfor- mance. In this work, we analysed the musical expressivity of jazz guitar and piano from two different perspectives: solo and ensemble performance. The aim of this paper is to study the influence of piano accompaniment into the per- formance of a guitar melody and vice versa...

yeah, forget the hype and hyperbolic marketing of all things A.I.

 

There is actual work being done to understand and interpret musical expression which is valid and creatively interesting.

https://archives.ismir.net/ismir2016/paper/000162.pdf

 

And here is a video with out the hype but plenty of excitement from actual musicians exploring “math”, sure its perhaps basic but is already “old” and does show a way forward…

 

 

PEACE

_
_
_

 

 

[jazzpiano88]

The large language models have received much of the hype lately via the media and ChatGPT.  These models will generate bullshit (completely incorrect answers or code) at the expense of accuracy because the reward from the objective function is not set properly.   You might say they don’t have proper error handling because they will give a bad result rather than an error condition (e.g. NaN). 
 

Where the real interesting work going on in AI is Generative Models.   I’ve worked on these for a living for almost 10 years.  They started out with a groundbreaking paper from Ian Goodfellow on GANs (generative adversarial networks) and have rapidly been extended, enhanced and repurposed to generate synthetic real world data.  Standard auto encoders can work as well but they usually don’t reproduce the random variations and look more synthetic to a human.  You generate these data from inputs that are not in the form of the output.   E.g. generate real looking images from text (label to image) or stick figures (image to image transformations).  These can be extended or applied music (waveforms), not just images (2d matrices of pixels) .     
 

One reason they work so well (at least for images) is that there is a two network ping pong game going on to train the network to both minimize the L2 error between real and synthetic, and maximize the confusion of a discriminator comparing the real and synthetic data.   These work well enough to completely fool a human observer I.e. you can’t tell if it’s real or synthetic and if you try, your accuracy will be ~50% — no better than guessing.  
 

What’s my point?  While this music doesn’t currently sound very good, it’s early in the game and results will improve.  With music you have a couple of things to get correct.  1. The short-term sound or timbre of not just one instrument, but multiple instruments, and (2) evolution of the sound over time (composition).  That doesn’t seem easy or straightforward to train.   I’ve thought about doing something simpler — a melody and harmony generator without worrying about timbre.  Once you have that you can midi-fy it into a VI package and add articulations.   It’s definitely a wide open space out there to do something new.  If you have a server and some GPUs the sky is the limit IMO.   Take some courses on AI.  Learn Python and tensorflow.   It’s not that hard and most papers now come with code you can use as a reference.  

[ElmerJFudd]

It seems to work the same way as AI image creation.  Which is to grab bits and pieces of all the digital images on the web and fabricate a conglomerate that meets the query in your request.   It often generates a useful image, but also some bizarre stuff  with errors like mangled hands, robotic facial expressions, etc.   The more specific and complex your need, the worse it performs.  

 

Our ears are more sensitive than our eyes, I feel, well at least a musicians’ ears.   And it’s very obvious the generated recording is a mish mosh of existing recordings with next to zero understanding of the language of music.  The more complex the style, the more it fails.  
 

IMHO what they should be doing with AI music is having it study a body of knowledge like Band-in-a-Box and use midi and virtual instruments to create the product.  Not tear apart completed arrangements from digital recordings and put them back together like Frankenstein.  

[stoken6]

3 hours ago, Stefan011 said:

Swedish pseudo jazz easy listening

How can Spotify tell Swedish pseudo jazz easy listening from other types of pseudo jazz easy listening?

2 hours ago, tapes said:

don't knock them Swedes - a lot of great jazz came from there, and the Nordics/Scandinavia

Definitely. NHØP (albeit Danish) springs to mind.

 

Cheers, Mike.

 

 

[jazzpiano88]

Results are only as good as its training.  Which is why generatives and predictions are so strange when presented with new input data.  Ais don’t create, they essentially interpolate using the weights established from training data.  Extrapolation (reasoning) works horribly unlike human intelligence.  

[Adan]

Unlistenable. I don't know who the narrator is, and don't care to find out, but he reminds me of the people I'd randomly run into living in San Franciso in the 2010's who'd make me think "what the hell happened to this City?"

 

I get that it's early days and this might be a stepping stone to something more interesting.  I'll wait.

 

Edit: after thinking about it some more, "unlistenable" is too harsh. "uninteresting" more accurate. Makes me think of when there's a jazz band playing in a dream and in your dream mind you know it's a jazz band, and the music sounds vaguely jazzy, but it's also way off because of course your unconscious mind assembles things weirdly.

[Stefan011]

33 minutes ago, stoken6 said:

How can Spotify tell Swedish pseudo jazz easy listening from other types of pseudo jazz easy listening?

Definitely. NHØP (albeit Danish) springs to mind.

 

Cheers, Mike.

 

 

Sorry, I was not knocking them down! I ment it literally. There is a fairly anonymous Swedish jazz band that Spotify (swedish company) uses for "jazz for something playlists". There's a video about it somewhere on Youtube. I love Swedish jazz 🙂  

[Stefan011]

45 minutes ago, stoken6 said:

How can Spotify tell Swedish pseudo jazz easy listening from other types of pseudo jazz easy listening?

Definitely. NHØP (albeit Danish) springs to mind.

 

Cheers, Mike.

 

 

Haha, It can not (probably). It was recorded for Spotify so they can make more money. Try seeking artists in something like "Jazz in the background" playlist. They can play, it is listenable. But I think properly trained AI can get there in few years. 

[Reezekeys]

For full-time musicians slogging it out in the "low to mid price category" 🙂 – i.e. most kinds of commercial music production, this is probably the real beginning of the end imo. I think background music, library music, short films, TV bumpers, etc. will all be done by AI. Bar bands are safe and those that tour with Beyonce or Taylor are safe. Everyone else in the middle - I hear Walmart is hiring!

 

In this particular case, some of us here have decades of listening to Bill Evans to compare this crapola to - so it's easy to dismiss. I'm actually not worried that AI will replace talented jazz players - imo the better AI jazz gets, the creepier it will sound (to me anyway)! No, it's probably the smooth jazz you hear in the dentist's office or elevators that AI will take over - maybe that's for the better! 🙂 

[Threadslayer]

5 hours ago, stoken6 said:

...concept of short/medium/long-term progression: call/response phrases,  tension/release, verse/chorus (or movement, whatever) structure, etc.

This ^ sounds like the beginning of a training set for AI to learn

[DJkeys]

Artificial Intelligence is not intelligent. It cannot really "create" anything. The algorithms are programmed by human beings with millions of if/then statements. AI will never replace real musicians.

 

-dj

[marino]

[ksoper]

Much of the "New Age" noodling we heard in the 80s sounded, to my ear, like it was machine generated. This isn't any worse. 

[ElmerJFudd]

11 minutes ago, ksoper said:

Much of the "New Age" noodling we heard in the 80s sounded, to my ear, like it was machine generated. This isn't any worse. 

It’s interesting how we can differ so strongly in our reaction to this  AI’s attempt at creating a Bill Evan’s track.  I think it’s largely to do with expectations and familiarity with the source material?  

[jazzpiano88]

44 minutes ago, DJkeys said:

Artificial Intelligence is not intelligent. It cannot really "create" anything. The algorithms are programmed by human beings with millions of if/then statements. AI will never replace real musicians.

 

-dj

Here’s an excellent book on the subject if you care to find out how it actually works. 
 

https://www.deeplearningbook.org

[KenElevenShadows]

I'm impressed at how far AI (really, more "machine learning" since "AI" is such a buzzword) has come, especially in the last couple of years. That it can even produce anything like this is technologically astounding.

 

I don't particularly care for the music at 8:09 personally, but I'm impressed that it can create something that is reasonably cohesive and readily identifiable as something resembling jazz.

[ProfD]

2 hours ago, Reezekeys said:

For full-time musicians slogging it out in the "low to mid price category" 🙂 – i.e. most kinds of comercial music production, this is probably the real beginning of the end imo. I think background music, library music, short films, TV bumpers, etc. will all be done by AI. Bar bands are safe and those that tour with Beyonce or Taylor are safe. Everyone else in the middle - I hear Walmart is hiring!

 

I'm actually not worried that AI will replace talented jazz players...

 

No, it's probably the smooth jazz you hear in the dentist's office or elevators that AI will take over - maybe that's for the better! 🙂 

Agreed. 

 

Everything written IMO...

 

AI will put folks out of work when it comes to creating generic music. 

 

AI won't  produce anything too much different from the music we hear in KB demos and arrangers.  

 

AI will never be creative enough to produce the next greatest musician the world has ever seen and heard.  

 

AI will never be a threat to musicians who are some combination of gifted, talented and creative.😎

[Docbop]

Video is an area where AI generated video is getting to be amazing looking maybe ever scary how real it looks.   The one recently getting talk is 

Sora below is a video on it.   Just saw a video on another tool called DScript for that edits video based on text.   The video of a YouTube creator did was so cool how quickly it edited his raw video from the text of what he was talking about and he could quickly add or change what DScript had done to make it exactly how he wanted.    AI is doing amazing things in the video world. 

 

 

[ElmerJFudd]

1 hour ago, Docbop said:

Video is an area where AI generated video is getting to be amazing looking maybe ever scary how real it looks.   The one recently getting talk is 

Sora below is a video on it.   Just saw a video on another tool called DScript for that edits video based on text.   The video of a YouTube creator did was so cool how quickly it edited his raw video from the text of what he was talking about and he could quickly add or change what DScript had done to make it exactly how he wanted.    AI is doing amazing things in the video world. 

 

 

Ya graphic art, animation and writing, music as well (especially for the non discerning consumer).  Definitely a tool for creative people.  Replacement for human creators?  Perhaps around the corner… maybe closer than we think.  

[JoJoB3]

1 hour ago, Docbop said:

Video is an area where AI generated video is getting to be amazing looking maybe ever scary how real it looks.   The one recently getting talk is 

Sora below is a video on it.   Just saw a video on another tool called DScript for that edits video based on text.   The video of a YouTube creator did was so cool how quickly it edited his raw video from the text of what he was talking about and he could quickly add or change what DScript had done to make it exactly how he wanted.    AI is doing amazing things in the video world. 

 

 

However story is always #1 and they'll never stop replying  "practical is better" (thankfully).

+1 for 'humans'