#195 - 23rd October 2024 - StudioSzene Day 2

I did not build a re-amp box. The DIY kit would have been 50€ and I figured I would not reeeeally need it.

Another full day…

… with the most important and relevant session for me being „production of radio ready vocals with AI“.

Kalli Reinhardt has shown what I knew and expected but never tried: Hyper produced pop vocals that sound a bit artificial anyway can be created with existing AI models with a level of quality that nobody would recognize in a negative way. In fact, he did it in various occasions and even got singer credits for creating an authentic 90‘s MC style shouting that is essentially his vocal idea, reproduced with five different models and made lo-fi. That‘s nuts.

He showed different use cases for using AI vocals and even trained his own model based on German singer Lena who he regularly works with and happens to have a couple of hours raw vocal material from.

Which is exactly what you need to train your own model.

An interesting learning and in a way to be expected is that these models work very much like guitar amplifier modeling. You get a snapshot of a specific setting that can be reproduced. But the clean sound snapshot will not convincingly produce a high gain sound.

The vocal snapshot in the AI model will reproduce a good chest voice if trained with it. But it will struggle with switching to the upper ranges, growling, screaming or whispering. For these changes in style, a new model would have to be trained.

At least for now. Kemper also brought the liquid profiles.