Originally Posted by Bass Thumper
I think you are cross-confusing huge general-purpose LLMs that are running on giant server farms (which is one specific implementation only) with other AI implementations, some of which are highly specific and have much smaller footprints. See above for examples. I have personally written AI/ANN algorithms that will easily fit on a small thumb drive.
NNs can be quite small - SynthV is a case in point, as are other NNs that resynthesize vocals and instruments.

So you're right - it's possible to do this and fit into a small device.

But with the ubiquity of cloud services, I'm having trouble understanding why a company would want to take that route. Call it a lack of vision, but given the option between locking users into a subscription model vs trying to sell a more limited model, I think the first option has a lot more appeal.

Even for the use case where a user wants a backing track, it would make more sense (to me) to have this performed in the cloud (via AI), and then for that rendered backing track be delivered in sections (intro/verse/ending/etc.) that could be switched to live during performance.

By way of analogy, think of BiaB RealTracks. An AI could generate new RTs dynamically, or it could generate them in the cloud and deliver them to you to use locally. The option to deliver them after building them in the cloud means that PG Music would have the cost of maintaining the infrastructure in the cloud, and you'd get the part you care about: the backing tracks.

Since the RTs or backing tracks are the final product, there's little compelling reason to have them generated locally - as long as it can be seamless.

So it's possible, but I suspect that's not the route it'll go.

-- David Cuny
My virtual singer development blog

Vocal control, you say. Never heard of it. Is that some kind of ProTools thing?