Yes, it’s Handel’s use of the timpani, creating that dramatic effect of “the enemies” being swamped by the water. This recording depicts that quite adequately by the singers being drowned out by the mad timpani player (probably using slightly more modern instruments).
Me too. But not faffing around with DAWs. Still, each to his own.
If you focus on the violin (viola?) performance, it’s above anything I’ve heard come out of Dorico, or a VST, so I would disagree with you there.
So if that’s where things are heading, and given more control over the details, the future’s looking rosy.
(But if, as you imply, there is an inherent limitation in the amount of fine-grained control AI can provide, then yes, perhaps this is a dead end.)
When this thread started it seemed to be another AI enthusiasts v AI-not-in-my-backyard-ists! In the last number of hours it has turned to a more thought provoking discussion which I appreciate. I am not one who is in favor of AI. The current model is to feed it a lot if indiscriminate stuff and it returns, spits out, or vomits stuff back to us. We, in turn find it marvelous or disgusting. (It couldn’t care less!)
But I think there could be some use for creators as well. We study the work of the Masters in order to learn the craft and tools of composing. We have teachers and instructors to guide us – not saying that AI should take that role! We use reference materials that help us with style, orchestration, harmony, arranging, etc. But these teachers and reference materials offer us ideas which AI could also do – in abundance - providing we realize that there are good and bad in the mix.
Currently, something like Suno takes our ideas and gives us its “idea” of how it should be; I don’t want a collaboration by some algorithm created by someone who has no idea of what I’m doing! But I’m not adverse to having an abundance of reference material that I can filter for style, harmonization, orchestral color, etc., or a “partner” that could quickly show me how a new idea would sound.
Thinking about what I “learned” over the years: in Schubert’s day it was generally thought that oboe and clarinet should not be paired in unison together. Yet it is wonderful in his 8th symphony! Or Wagner’s harmonizations – things that no one else had done would not have been in an AI’s dataset. There is still plenty of room for creation.
On the other hand, there is plenty of stuff that masquerades as art available at your local dollar store. Humans are just as capable as AI to create sub-par artworks!
— Jim
That may be true, I don’t have very much knowledge of VSTs at all. To me it sounds similar to the results achieved using a VST by an experienced user, but from the limited experiments I’ve done you would probably have to interpret and shape VST output using a MIDI controller, whereas obviously this generates a result without requiring one.
That said, these AI tools do not appear to read pitch and rhythm data in the way something like Cantai or NotePerformer does. I think Cantai is generative AI but NP isn’t. So what you would be looking for is something more like a Cantai for instruments, where you input notation data in some form and it is broken down into parameters and rendered into audio (presumably—I don’t have knowledge of compsci either). At the moment this is computationally expensive.
Another more promising possibility would be an AI that generates performances for VSTs, i.e., it produces and outputs all the automation data and can update it based on prompts. This may already exist somewhere. It could also be combined with non-generative/algorithmic sound modeling so that you can do this without having to load 600 GB of sample libraries.
This sounds a little like one asking for NotePerformer’s now defunct NPPE.
NPPE and NotePerformer are not AI of any sort.
I would like to see more detailed / musical shaping in Dorico especially for long sustained notes. AI is potentially one way of doing that. I know users working in DAWs have also asked for similar capabilities. The AI would have to have a good understanding of each particular library though in order to be highly effective.
No question of that, but Naomi’s last paragraph seems to be asking for NP/NPPE capabilities, which can be realized even without AI.
This sounds very interesting but for me not yet ready for prime time, so not worth the investment of my time. But it is definitely something to keep an eye on for my orchestral work.
The ability to compose in Dorico, export audio, and render through an app like this, if and when it is ready for prime time, is very appealing. I feel like the OP, I like the idea of being freed from the preoccupation with the complex technical underpinnings of a DAW.
Still, I am not a professional delivering compositions for live performance or audience consumption, and that affects my viewpoint. I just want to hear my unique arrangements and compositions played back in the most convincing way possible.
Essentially yes, although I understand why NPPE was discontinued (a large amount of work and effort, for results that provide improvement in sound quality but are still inferior to NP on the interpretive level, and lead to a more adversarial relationship with the companies that provide sample libraries). The only thing AI could bring to that is convenience via natural language processing, i.e., instead of trying to draw a realistic amplitude/frequency of vibrato, tempo variation, dynamic variation, etc. for each individual note or phrase, or adjust rules and symbolic logic to try to get the results you want, you select 20 bars and prompt “the violins should use more vibrato here”, “the soloist should play this section with rubato”, “this theme should have a more cantabile quality”, with the output being the automation data.
And loads of reverb.
There are passages in the bass line that are clearly not human.
But the whole concept underlying most AIs is to produce what has been identified as “desirable” by the trainers. Often that results in pablum, and that often is sufficient.
I do agree with Pete’s premise (which I believe is) that this will hit the makers of the expensive VST libraries very hard. I always considered a $2000 price tag outrageous anyway. These companies will disappear.
So I gave it a try today. The quality of the sound compared to say, a VST, out of the box, is indeed impressive.
But it took so many creative liberties. In the end, while the result sounded “good,” it was completely useless because the transformations were too great. I can’t imagine relinquishing this much control in any professional music setting, but I’m sure it will be a fun toy for people to noodle around.
I don’t think Suno is really going after the market of composers who use those expensive VSTs, for the reasons said above – it changes your work too much and adds all kinds of random things. It may sound good but if you make music professionally, and collaborate with clients, that’s basically useless.
What I do think, is the smarter companies in the VST library space will adapt and leverage this technology. I’m sure it requires a lot of financing and resources to do so - so probably only the biggest ones will have the capital to invest in such an undertaking. I’m sure indie Kontakt developers making generic orchestral libraries will become irrelevant - but those who create unique timbral/texture libraries will likely survive as they can offer something AI doesn’t do. I’m sure the entire space will change in the coming years, but it seems that major AI developments are not focused on professional creative use.
I wasn’t referring to Suno in particular. If Suno can produce results comparable to the $2000 packages, then somebody will package that as a VST. Remember that many of these AIs are thiefware, in that their contents are all stolen and the AI models themselves are shareware.
I know you weren’t, but I’m keeping in mind that Suno is reported to be up for a valuation of $5 billion, and one of the biggest VST makers on the market, Spitfire, was recently acquired by Splice for $50 million - which is peanuts by comparison.
My point is, thieving aside (which I totally agree), they have the money and resources to build and scale - but they aren’t going to be focused on niche professional markets like composers. That’s not where the ROI is. So to me it’s a little bit of an apples to oranges comparison. Yes, Suno sounds better than many VSTs, but are there developers in the VST arena who have the ability to develop something as impressive and at speed, but can also be legitimately useable for professional productions? That’s where I have my doubts.
Here is the result of my Suno experiment. First of all what I fed it was 30 seconds and honestly I recognize nearly nothing that I originally wrote - so accuracy of composition is way off. I know there’s sliders to adjust that and I could prompt all day long, but I didn’t want to pay the asking price.
I’m sure the quality will improve but it sounds smashed to bits: the artifacts, noise, distortion, and over-compression. Also when the violin goes super high it starts to get quite synthy / VST-like haha. But it was fun to play around and see what it’s all about.
Yes, I agree Suno has bigger fish to fry. But there are literally thousands of open source LLMs now. Here is a list of a tiny fraction of them:
At some point, there will be open source models that can render sound at the same quality as Suno. If somebody can train one of those to respond to MIDI or to the data flow through the VST3 interface, that will very quickly eliminate the expensive sound libraries.
Keep in mind that Spectralayers and Cubase (at least I believe this is the case) both utilize several open source AIs to accomplish their unmixing. It seems inevitable that something similar will emerge as a VSTi.
Back to Pete’s main thesis. It is a matter of speculation if/when this will happen. But in this particular neck of the AI woods (sound rendering), I don’t find that much more dystopian than the sound libraries themselves or NotePerformer. Arguably, they can all put humans out of work. By the same token, Dorico (Sibelius, Musescore et al) puts plate engravers out of business. But they were out of business thirty years ago, with no push from AI.
A certain pace of technology progress is expected and accepted as part of social evolution. I think the main issue with AI is how suddenly it has come on. Even today, there are few clear cases where large numbers of humans have been replaced by AI. Somewhere between 70-90% of businesses who have pursued an AI strategy have so far, failed to achieve any meaningful savings. But everyone expects that is just a matter of time.
One version of this story predicts that in a decade or two, there will be very few jobs that robotics + AI can’t do better than humans, and without a completely new concept of society, our civilization will collapse. Another version of the story says that the AI hype is overblown. Robotics + AI may take over a great many tasks that nobody wants to do; repetitive, boring, detailed things. But humans will find ways to apply themselves higher in the value chain.
Historically, one of the outlets for human ingenuity, when technology freed up the workforce, was art. Architecture, public works, graphics art, theater, music. These were all hallmarks of an advancing society. What is different this time is that AI seems to be attacking art directly.
Interesting - there is kind of a hissing sound in the background. Also the melody seems to take a bunch of illogical turns.
The one thing I keep coming back to regarding AI and music is that to my limited knowledge on the subject an AI cannot “hear” with the senses the way humans can. Sure, it can measure wave forms, read pitches, read timing, apply adjustments based on algorithms, anticipate articulations and so forth. They can even imitate a vast array of imported compositions. But it cannot hear with the senses.
Noteperformer does great for what it is, as I’m sure Suno will be. But people, their sensory perceptions and their responses to them, these are the point and purpose of music.
So on this basis I go along with those who say AI music software can’t replace the part of music that is essentially human and personal. That said, I’m equally as sure that properly implemented it can be a valuable assistant that saves the composer time, provides the opportunity for give and take that in the end can sometimes stimulate ideas the composer may not have otherwise had.
Bring real value, yes. Replace some jobs, yes. Replace humans, no.
I tend to agree with you. The problem with AI, as it currently exists, is that most AI has no ethical guardrails to prevent unauthorized use of copyrighted materials. Depending upon the training environment, it is possible that AI entities could train on the same existing music and similarly impose that training on all of its future work causing multiple instances of copyright infringement.
Probably the only attempt of stating the use of ethical AI (to my limited knowledge base) is Richard deCosta’s work with Cantai. Using “original” voices and providing those voices (or groups of voices) a share in his profits, Richard has created a basic guardrail for his work. As the product evolves, it may well need additional ethical guardrails but that will become apparent during the evolutionary process.
I do not worry about AI so much as a tool for composers and musicians but more as a uncontrolled entity in social media and other online environments with the intent to defraud and/or cause harm. I suppose we could easily see that in the world of musical AI, too, yet most music composes and practicing musicians involved with AI want their creations as much as everyone else’s to be used within the framework of copyrights.
There appears to be multiple facets of AI that we do not yet understand fully and that do not necessarily have ethical guardrails.