Is there any news on what Cubase is going to about AI integration? I’m seeing a lot of AI production tools coming out recently with varying levels of interaction. Some seem to be stepping on the toes of traditional DAWs.
It seems to me if Cubase doesn’t get on board in some way it’s going to become a relic very fast.
Personally here are a few things I would like to see coming to Cubase:
Using automation of inserts the user has added allow AI to create a mix that can be tweaked / edited as normal for the whole song or a region
Have AI restore badly recorded audio or audio with glitches in it so it is production quality.
Use AI to clean up all automation in a session i.e when automation is copied around it creates a lot of mess at the moment
I’m on the fence about AI sound generation but do think there is a place for it. Like if I have a near completed track AI would be useful to add in specific sounds based on the context of the underlying audio. As long is it producing isolated sounds I think AI generation is probably ok.
Anyway I hope Cubase is not taking some moral stance on this as AI could save a lot of time in music production without harming creativity.
There have been loads of threads on this forum discussing people’s ideas, and/or resistance to ideas, on this general topic. However, with respect to “news”, Steinberg generally doesn’t announce what is coming until it actually arrives, be it as a new feature in a maintenance release or as part of a major paid update.
That said, in Steinberg’s annual Cubase user survey a few months ago, there sure were a lot of questions regarding various areas of potential AI, so it would at least seem they are considering potential directions on this front. Given the timing of that survey, and the most typical timing of a major update coming in November, I’d be inclined to guess they were asking with regard to a Cubase 17 (or beyond) instead of a Cubase 16 (based on what I’d guess to be development cycle times for major new features plus alpha/beta testing time on such new features). But who knows – maybe they already had some experimental efforts in progress and just wanted to get some input on which ones to prioritize for finalizing in time for the next release?
If I had a buck for every poster on here who proclaims their better-than-Steinberg knowledge of the DAW market I’d be doing pretty well by now. It’s astonishing how Steinberg manages to stay in business when all these folks know so much better.
Thanks, I was wondering if there was any news, announcements or dialog going on. With this being such a controversial topic I thought maybe Steinberg would have engaged with the customer base, but it appears not.
I was asking because you had mentioned welcoming a “moral stance” on AI, and implying (I think) that you felt AI harmed creativity.
I should have been more clear, apologies - I was wondering if you feel that way about all AI, including the AI currently in Cubase (SpectraLayers Go), or just the AI of the sort that newspeak101 mentioned in his OP?
Basically, I’m wondering if I am missing a nuance here that you appreciate.
Well, in a very important sense (i.e. influencing future Cubase directions), Steinberg’s inclusion of all the AI-related questions in their annual customer survey was engaging with the customer base on this topic. I don’t recall specific questions at this point (I’m thinking it was 2-3 months ago), but it asked questions that referenced a whole bunch of potential areas where AI could conceivably be used, from the generative stuff so many think about when the term “AI” is used in the context of music production to various project organizational possibilities to mixing assistance. I think there was also the typical sort of multiple choice numerical ranking type things where they might ask about a specific category of feature and ask you to rate it on a scale that went from equivalents of high priority to “don’t waste time on it”.
Assumedly, they also take some pulse on repeated discussions within these forums, even if they don’t tip their hand to what might be coming in future releases (or even general directions on where things might be going).
Or they look for ideas Imo, it’s all down to logistics.
I have a pretty extensive daily set of routines with ai and imo, I’ll tell ya, ai stands for “artificially inclusive” before anything else. Ya gotta define an area of “what is the focus here?”
Whatever the underlying learn codeset…it has to be continually tweaked and updated by the “codester”. Logistics…
Ok, a focus. Steinberg decides to integrate a realtime ai “learn routine” into an old-fashioned match-eq (for example) as a baby step. You activate the “I-Is-your-magic-eq” button on a cubendo mono channel of a snare drum or drumset or whatever. Click the panel, invoke the 'this becomes that" button. The “that” being the target “thing sound” based completely on the intricasies of the existing “this”…emanated from either your own personal “that” category or your text input. There are no extensive samples on your hard drive…no space hogs. It’s all realtime calculation. Result for the 5-minute track generated in less than 3 seconds. Instantly tweakable, instant relearn, instant undo. Conversation with the module can even be “no, I think I want that snare eq to be more like the one in the elevator shaft on 1968 “The Boxer” but no reverb…you have too much 8k on the result you just gave me”…(warning…internet connection required cuz “I-is-your-magic-eq” has to go online to check its database on “The Boxer” and that will take 2 nanoseconds)…and then voila…instant tweak. All within the Cubendo v28 track panel.
While the concept can be applied in zillions of areas, let’s say SB starts with the magic eq.
Does everyone understand how much time must be spent updating the underlying code learn on a daily/weekly basis?
Massive amount of time…imo. We’re not talking static code like an ancient typical “match eq”.
Logistics…time, manpower, resources. And if SB/Yamaha were to choose farming out some of this to the “holders of the brains” (anthropic etc) well..for that…they have to pay.
Which then jacks up the price of Cubendo.
You get these routines integrated into a bunch of daw module areas…you now have a whole lot of code guys on the payroll.
It’'s not like you can slap on a magic eq codeset learn routine today and then not maintain tweaking until “next year”.
Things are moving way way way too fast with learn code to adapt that kind of “periodic tweak to new breakthroughs”.
Absolutely. So many hear AI mentioned in the context of music and automatically think “Suno” (or some equivalent generative AI that can set your lyrics to music, write an entire song for you, produce a recording of a song you’ve written, or some combination thereof). But many of us are already using some degree of AI in our music productions in forms like the iZotope “assistants” (i.e. master assistant in Ozone 10 or later, or maybe even starting earlier, similar assistant in Neutron 5, reverb assistants in their reverbs that deal with masking of the input signal, RX’s assistant), Waves plugins that create initial settings based on analyzing your signal (I think most of the Curves products, their Vx plugins for de-reverb and noise reduction, the StudioVerse Effects function that looks for presets based on your input signal, etc.), SpectraLayers’ stem extraction, etc.
I wonder, in the case of your example, how much is tweaking the underlying code set and how much is tweaking/updating the training data. I’d be inclined to think the code would start to settle over time (at least other than when new major developments in how to look at training data occur), but the training data would continually need to be augmented and modernized. This is an area, though, that I think would be best left to the specialists, as opposed to a DAW developer. I like Ozone’s Master Assistant, and I’ve used it ever since it first arrived, though, in that first release I never ended up using its results. Somewhere along the line (in or prior to Ozone 11 since I haven’t upgraded to Ozone 12), though, it started getting results that were much closer to what I was trying to achieve, be it through Ozone or other mastering-oriented plugins I was using at the time, and there have even been times when I’ve used its results directly, alongside other cases where I’ve made relatively minor tweaks to its results. Those who don’t want that sort of thing simply don’t pay the extra money they’d need to put down to get a plugin like this.
This text input notion is something I’m personally not expecting, and somewhat hesitant about even wanting. Some intelligent tweakability (e.g. using specific tags to refine suggested preset results in a StudioVerse Effects) is one thing, but that’s not really AI, but database – the StudioVerse AI thing is the part that analyzes your signal to figure out what tags it should search for in their database of presets. Of course, anytime anything is analyzing a signal to suggest presets, there has to be some training database somewhere that contributes to what suggestions are made, even if that logic is frozen into code, rather than something that has to ask the database each time.
Beyond the language consideration, this is where we potentially get into dicey legal territory, in particular for this example the reference to a recording under copyright protection. It’s cool to potentially have the functionality, but it’s also critical that any such training data has been licensed for that use.
Your internet connection must be WAY faster than my (pretty darn fast) connection. The Celemony Tonalic whose integration was prominently featured in the Cubase 15.0.30 announcement requires an internet connection. So does Waves StudioVerse if you’re accessing their online preset database versus just your own local presets. On the other hand the iZotope assistants and Waves Vx and Curves stuff do their jobs locally. I guess the point is scope matters immensely, especially once you start taking into account data and/or content that is outside your system.
My guess is they take on this type of project by licensing it from someone with more specialty in the specific area needed and/or by using open source code. Of course, they also need to integrate it in with their Magic EQ Plugin’s interface. (And I’m still skeptical about the natural language interface need.)
“Rolling their own” (i.e. from scratch) wouldn’t seem to me to make much sense, especially given there are already plenty of products (at least if we ignore the natural language interface and online database to be searched from inside the DAW aspects) out there that already do this sort of thing (Waves Curves AQ especially comes to mind, and probably also the Neutron 5 track assistant – I didn’t upgrade to V5, though, between upgrade price and really not using Neutron that much in general).
On the other hand, if Cubase were to add some intelligence in areas like organizing your projects, updating batch export facilities in ways that avoid the multiple batches that are needed nowadays, automating project archival, etc., those are things that have to be done at the Cubase level, and that they should already have some degree of expertise on (be it directly or from various user inputs they’ve gathered over time). And some of that level of needs won’t go away. (Sadly, the additional batch export needs were voiced by multiple people, including me, on these forums back when the facility was first introduced, but there have been no further developments on that front since – multiple major Cubase versions later.)
BTW, just to add one extremely timely (due to discussions in another forum thread) idea on a Cubase AI improvement: A facility to analyze your project, including its bus structure, sends, etc., and give suggestions that would help Cubase better allocate the load across whatever CPU cores your system has – i.e. what’s impeding being able to parallelize it more? – could be quite useful. The ideas in the Performance Monitor might be helpful to some users, but I’ve never found them to tell me something I didn’t already know (e.g. freeze this track) and probably would have already done had I not needed it to be how it was at that moment (e.g. because I’m working on tweaking the inserts for the track). On the other hand, if it told me that some specific sends that affected multiple submixes was forcing all those submixes into the same CPU core (not necessarily in that language, of course), that would be valuable information that could give me some ideas for revisiting my project’s bus structure.
.
Imo, it’s all down to logistics.