#but the plus banks are actually AI models trained off of their concatenative samples iirc | Explore Tumblr posts and blogs

bmpmp3 · 22 hours ago

Text

this is another thing that probably doesnt matter at all but as someone who's interest in vocal synthesis is in large part because of the software and technological aspects, every time i see someone trying to explain the use of deep learning/neural/AI/etc in vocal synthesizers and they say that "the only thing the AI does is help make the pitch transitions smoother" im like white knuckle gripping the table muttering under my breath like no....that is. incorrect.

#there is a big misconception that deep learning synths technologically are the same as concatenative like a series of samples #stretched and stitched and resynthesized together with the 'AI' only referring to an automatic pitch system #and i understand where the misconception comes from. its probably a combination of early marketing of deep learning synths #(am i insane or did ahsoft use to market AI rikka etc as standing for 'automatic intonation'.... did i make that up)#plus trying to separate ai vocal synths from like chatgpt and whatever #BUT. that is not how it works. i think the only synth ive seen that does have that functionality is the very recently released miku nt2?#which i think is still in beta anyway LOL #i thought there was maybe some early synthv banks like the plus banks that did that too initially #but the plus banks are actually AI models trained off of their concatenative samples iirc #but yeah.......... ai voicebanks are just straight up deep learning models of voices with a lot of built in control tools in software #(what notes to sing what parameters to change tone etc)#the vocal provider sings a whole lot. the programmers go in and carefully label all the data. etc etc #they are more ethical than like some of those sketchy song generators in that the data used to train these models is obtained via #licensing and direct input by vocal providers who are getting paid and giving consent etc. but the technology is the same type of thing #i dont even like or care for randomly generated gpt whatever the fuck i find it super uninteresting 99% of the time #but i do love a good ethically made deep learning based vocal synthesizer voicebank and i really dislike technological misinformation #dont stand to close to me or i will start explaining to you about linear predictive coding speech analysis. DO NOT test me

5 notes · View notes