Two questions for RC 6 already

[tatanka]

Strange, seems as definitely not one of the regular guys but haveing the first post on RC 6

Before the questions maybe one answer too "why" (not one of the regular guys)

If it would not have been mentioned in the video about the Neural phrase engine gladly how should any not long time freak realize that the download that is needed for using the really groundbreaking new ENGINE (far more then a feature...) is under:

Settings (that makes sense) but then "Miscellaneous" (so all that is NOT important from my understanding....) and ok, if you find you way there: AI

Now my questions:

- first: the generated tracks are all on the same chord, no matter how long the generation is. So maybe for the noobs like me or the totally new guys coming for this feature: how to make this with a chord progession (generate tracks following a chord progression) ? I am sure 99 % of you guys here in the mc forum new, but for the rest a video showing this would be really helpful

- second: how does the new Neural phrase engine could/should/can work together with the idea tool ? Video showing this requested too

One thing I really want to mention: please keept the videos coming on this youtube channel and in the (very good ) style of this whats new neural phrase engine.

One place to find them all.

In a way evrybody can follow step by step

[tatanka]

Ok, it seems the new youtube video page really opens up the so often missed way to really get into it. For me this is a bigger update then from RC 5 to RC 6 simply cause I now HAVE a way and a PLACE to get at least into the most important stuff in a way I can get.

So I could have answered myself the question about adding chord progression to the generated track myself with the appropriate video.

But that leaves a follow up request (which might make sense or not, but when you want to start from a harmonic idea....): that is an option in the neural phrase engine to keep the chord progression/chord track for the new (even for create totally new) creation.

The question about using neural phrase engine and idea tool together still open and definitely needs a video on the youtube channel cause I coule imagine bringing these two together will could really be it....

Little addon: I really wonder about the little reaction on RC 6 in here, in kvr and in general. I mean the neural phrase engine brings AI to YOUR PC, no subscription needed, nothing. And I am really amazed of the bandwith of the tracks it creates. Its not that you get the number 1 hits instantly galore but from country/folk songs to classic piano compositions on a CLICK, this is so amzing but seems not much people even realize that its there....

Not my part: but I would really focus on bringing this neural phras engine AND the new really starting youtube video site into attention...

[kottke]

Strange, seems as definitely not one of the regular guys but haveing the first post on RC 6

Before the questions maybe one answer too "why" (not one of the regular guys)

If it would not have been mentioned in the video about the Neural phrase engine gladly how should any not long time freak realize that the download that is needed for using the really groundbreaking new ENGINE (far more then a feature...) is under:

Settings (that makes sense) but then "Miscellaneous" (so all that is NOT important from my understanding....) and ok, if you find you way there: AI

Now my questions:

- first: the generated tracks are all on the same chord, no matter how long the generation is. So maybe for the noobs like me or the totally new guys coming for this feature: how to make this with a chord progession (generate tracks following a chord progression) ? I am sure 99 % of you guys here in the mc forum new, but for the rest a video showing this would be really helpful

- second: how does the new Neural phrase engine could/should/can work together with the idea tool ? Video showing this requested too

One thing I really want to mention: please keept the videos coming on this youtube channel and in the (very good ) style of this whats new neural phrase engine.

One place to find them all.

In a way evrybody can follow step by step

I have the same questions about the Neural engine and agree about the Youtube channel videos.

I can see the power in the Neural engine, I just can't seem to harness it in a structured way as of yet.

[musicdevelopments]

Thanks for the questions!

There is still some work remained (a few items on my to-do list). At the moment the engine outputs 'absolute' notes only, that is why there is a single chord on the master track. It would be great, in the next version, if the notes are analysed, and chords are set up on the master track, with options to convert tracks to chord-relative or scale-relative form.

It is not yet integrated into the Idea Tool workflow, which would be very powerful. This will also come soon.

I am investigating other uses, or even training networks myself, but this will take time.

Thanks!
Attila

[tatanka]

One additional request which could be done in steps from "raw" to "fine"

To be able to choose a kind of "style" for the creation. For the first step for me it would be great to just being able to choose "classic" or "non classic". For later differentiationl of course: the more styles the better, from usual (like folk, pop, rock, country, ambient to ....)

But after having some nice different song ideas with first try (a nice jazzy one, a nice folky one....) lately I ended up with around 10 classical flute orchestras, string orchestras, piano orchestras.....and I am not into any kind of classical stuff at all.

So being able to excludes this would be very helpful as a starting point.

[musicdevelopments]

Hello tatanka,

you are right, that would be great. I think I will train different networks for different genres because the style cannot be set right now, although the track instrument selection influences the style.
When you find a good idea, you can make a small change in the temperature/top-p/top-k settings to get another variation of the same idea. If the change is big, you'll get another composition...

Thanks,
Attila

[BluGenes]

Hello tatanka,

you are right, that would be great. I think I will train different networks for different genres because the style cannot be set right now, although the track instrument selection influences the style.
When you find a good idea, you can make a small change in the temperature/top-p/top-k settings to get another variation of the same idea. If the change is big, you'll get another composition...

Thanks,
Attila

If you could, and I see this happen all the time (not just here in this example), the terms and functionality are lost because the terms are known by a select group of people, for example here, the top-p and top-k are no where explained (no hover help). It's the old problem of "a programmer writing for other programmers" explanation issue. Always made me go "grr" when encountering this pattern. lol.. I have seen this before with other software.. happens in haste getting the idea published, and not really completely thinking it completely through.. It may work, but, the user may not understand why it worked. I suppose I try to always consider how the user see's it.

[musicdevelopments]

So you recommend to dumb it down to a single setting?

They are of course explained in the hover help as well as in the user guide:

[BluGenes]

Hover over the top-p or top-k doesn't work. Your image is for the temperature...

[tatanka]

It does work for me (sometimes....) but it is cumbersome. Wait....wait....and if you are lucky it might appear.

I think that is how the whole help/info system in general works so it would be difficult to improve but of course I would prefer it to work instantly and all the time.

I dont know if possible but maybe on the long run a single place (like in Kontakt, Ableton) where all the info appears would be a more stable and better solution.

Great idea with the different models for different styles

[lovemusic15]

Hi Attila,


I'll be honest — I agree with this comment: [It does work for me (sometimes....) but it is cumbersome. Wait....wait....and if you are lucky it might appear.] To me it feels like playing a fruit machine. Even when you get something that resembles a song starter, it still needs a lot of work.
I understand you plan to train more data, but based on the current results, how will that reliably improve things? Genres and styles are complex — a “jazz-funk” label can mean very different things to different producers. Training only a few named genres won’t cover subgenres or hybrid styles that many of us make. Allowing users to train with their own data seems like the best path forward, since many producers have large collections of MIDI files and custom material they regularly work with.

Can you confirm whether user training will be possible? If not, explain how you’ll ensure future dataset updates deliver consistent, usable results across the wide variety of musical styles people actually use.

Thanks

[BluGenes]

My two cents is, it won't work like you think. In reality, it takes a very large amount of data as input, and almost equally large amount of what should be expected as output. That means it needs to know what the midi should look like when it is generated, before it is generated.

I have done tons of research on this for both myself and Atilla and the best possible way to do this is with a crap load of hardware, which you can get access to, but, would cost too. The process needs to run a system with a GPU, and not just one GPU, but hundreds of these.

You have to supply a data set too. It needs not only the genre and all, but, you have to supply the expected results.

OpenAI has great resources if you are interested in reading about the process and Google offers access to their hardware.

Just search on "model fine tuning" to get started.

[musicdevelopments]

Thanks for the feedback. I fixed the hover help today and made it more reliable so it always appears. FYI you can set the help delay under Settings → Miscellaneous.

[lovemusic15]

My two cents is, it won't work like you think. In reality, it takes a very large amount of data as input, and almost equally large amount of what should be expected as output. That means it needs to know what the midi should look like when it is generated, before it is generated.





I understand your concerns. My plan is to supply full arrangements and complete songs that contain all the information about how a song should look—how melodies, basslines, and riffs typically appear and relate to each other within a genre. Because I’ll supply material from one specific genre, the results should stay consistent; I suspect current models are trained on many genres, which is why outputs can feel random.

I’m not a programmer, so I may be missing some technical details, but in theory it works like this: with the Idea tool I load all my basslines and melodies and then hit randomize to generate new songs that feel like my style. It’s the same as when I fill Ableton’s Drum Rack with all the kicks and claps I like and hit randomize—I always get kits that sound like my taste. That’s why, as long as I know what information the neural network needs, I can supply it.

I appreciate your two cents—please let me know of any other limitations that may arise or what else we would need to build our personal dataset.

[musicdevelopments]

BluGenes is right, this is not that easy. To train such a network you need MANY files, a few thousands are not sufficient. It would be amazing if users could train neural networks for their own music, but this happens in python and involves multiple steps and takes very long. I have some secret plans with this though...

Thanks,
Attila