Originally Posted by Rustyspoon#
David,
Billy showed off his "local" LLM setup sometime ago. It was sort of in the back of my brain. A few days ago I've read on it, downloaded and briefly tested. While it is slower to respond, it produced viable answers, even on intricate questions. There are many free models. Some tailored for "reasoning" some for "coding". I am taking care of a friend, who needs a lot of attention, so not much time to play with these toys right now, but I definitely want to experiment with this FX plugin.
I've been re-doing my vocal synth program. Now that I've got the equivalent of a DSP expert on hand, I've been able to get past a number of hurdles.

I find that it works well when closely guided. I'm using the free version of Gemini, with occasional assists from the free Pro version when it can't track down a serious bug. I'll probably have to get a subscription, because I keep running into the issue of it having a limited number of tokens and so constantly losing context.

It can do common coding things quite well. For example, I'd been planning to use a spectral morphing algorithm of my own design to handle missing phoneme transitions - something that would require tagging phonemes to identify formants. Gemini suggested I use Line Spectral Frequencies instead - an algorithm that's been around since 1974 and used in phone transmission. It relied on LPC filters - something I'd never been able to implement correctly. Gemini free was able to crank out the functions I needed, but Gemini Pro was needed to debug it, where it immediately identified that Gemini free was trying to something nonsensical.

Between the two programs I eventually got it working. That's something I wouldn't be able to do on my own.

On the other hand, Gemini free is constantly hallucinating away key parts of the program and introducing regressions. By the time it's solved one problem, it's entirely forgotten what the rest of the program does.

Sounds like I may need to get a hold of Billy. smile


-- David Cuny

My virtual singer development blog
Vocal control, you say. Never heard of it. Is that some kind of ProTools thing?

BiaB 2025 | Windows 11 | Reaper | Way too many VSTis.