Life Beyond Tagging.... who else dreams about NI offering AI-assisted preset browsing?

nightjar

The new Komplete Kontrol 3 came out a few days ago, ushering in a new era and technical foundation for NI's virtual instrument catalog.

I'm sure there is PLENTY on the to-do list for the KK team in bringing priority features into the KK3 realm.

BUT... I hope there are some back-room plans getting things hashed out for "Life Beyond Tagging".

The AI technology from iZotope seems the perfect tool to bring a whole new method of finding the right preset for a song arrangement.

Imagine thousands of NKS audio previews being objectively analyzed for their spectral/time-based content, and then being assigned "soft tags" that are far less susceptible to the frequent near-meaningless whimsy and odd subjectiveness of current "hard tags".

And the ability of finding presets that are near equivalents of some short audio example you could drop in a search box... wow, that would be so handy and fun.

Hoping such dreams become reality over the next few years in KK! (Kontakt too of course)

nightjar

I was thinking about the problem of user input for preset searching in the upcoming age of AI-assistance.

It seems that QWERTY input is vital, but unfortunately a QWERTY keyboard is a cumbersome thing to put in the prime real estate of a keyboard controller.

So... imagine if all NKS2 presets have had their audio previews AI-analyzed and a soft-assigned a bunch of sonic quality values based on a table of gradient scales.

Then the 8 knobs of the S-MK3 keyboards could be assigned based this table of sonic qualities and roll thru the various gradient levels producing a listing of presets as you desired.

One such quality might be: Dull/Bright

Another might be: Organic/Synthetic

These two examples are a bit simplified... there would a few dozen assigned values, some being rather more sophisticated in their content analysis.

Such a search method might lessen the need for QWERTY input.

Kubrak

Keyboard could be well used as QWERTY keyboard. C might be A, D would be B, E C and so on. NI could sell nice stickers to mark keys.... They could be in new NI script.

There is no need for AI asistance to find similar samples/sounds. But yes, they might be tagged by certain metrics, maybe AI. But to train AI, one needs first to have huge training set. So, human would first decide if somerting is dull or bright or in between.... And so on....

Generally, I would not expect much from AI, it is just tranformation function from nD space to mD space.... If we speak about multilayer ANN.

nightjar

But once trained by a huge training set... AI and/or ML can be a super-fast tool going forward to analyze a MONSTROUSLY HUGE set.

MUCH less human labor/time/cost needed for the infinite amount of stuff to follow.. and it will continue to grow and refine.

DunedinDragon

https://community.native-instruments.com/discussion/comment/106051#Comment_106051

You're missing the main point. AI can't simulate human intelligence when that human intelligence can't possible define what is "too bright", "too muddy" with any kind of consistency across all people. Maybe over enough samples from someone using it, it could learn what YOU think is too bright, but that could easily also be influenced by the style of the music and your emotional reaction to it. AI isn't about programming it's about learning.

Kubrak

@nightjar

But one does not need to use AI to emloy automatic computer tagging. Classical clustering and so may be used. And concerning AI. It could have been used for that task for decades....

I have started using AI 30 years ago... But, I am not an AI expert by far....

nightjar

https://community.native-instruments.com/discussion/comment/106057#Comment_106057

A lot of the analysis/qualities would be purely objective numeric measurement… with the finished product, the user would make their subjective choice as they turn a knob while searching for desired presets.

nightjar

Think of all the qualities an image might have… an objective measurement of how much red, how dark/light and so on

apply that to audio previews