Artificial intelligence officially gets scary
- Thread starter Leo
- Start date
wektor
Well-known member
well there is https://www.riffusion.com/ as mentioned in the other thread, but still very far
wektor
Well-known member
already there, telling from the audio conference ive been to last year. live source separation is all over DJ apps, turns out your phone has a chip powerful enough to do things like that
ghost
Well-known member
Source separation is nice but I wonder about say, rewriting lyrics in real time, or doing style transfer to make the drums in one track sound like ones from another. Or bridging multiple tracks by synthesizing space between them. A ton you can do especially if you don't quite need studio quality.
I think both google & riffusion have been tested in very specific styles or tastes and as soon as they go away from a few narrow channels they just fail incredibly. Type in 1994 style jungle music in either and see what comes out, that music is nearly 30 years old and it doesnt have a clue. A long way off the auto-generated gear
That dj app is good. If they start implementing stuff like that functionality in a vst or DAWs direct then passing it into hands of people already making music can have a field day tbf
That dj app is good. If they start implementing stuff like that functionality in a vst or DAWs direct then passing it into hands of people already making music can have a field day tbf
shakahislop
Well-known member
the AI 'tiktok voice' seems to be totally accepted (within the medium anyway), despite sounding pretty rough. i was surprised by that. but then i'm surprised by most of the aesthetics of the online video form, like the jagged youtube jump-cut style and so on.
tiktok, insta reels and some corners of youtube seem really ripe for automated content i think. the audiences don't care, the quality is so low anyway, and the content you see is so algorithmically decided anyway, it wouldn't feel out of place in the medium.
tiktok, insta reels and some corners of youtube seem really ripe for automated content i think. the audiences don't care, the quality is so low anyway, and the content you see is so algorithmically decided anyway, it wouldn't feel out of place in the medium.
Clinamenic
Binary & Tweed
Mr. Tea
Let's Talk About Ceps
"broseidon"! amazing
Clinamenic
Binary & Tweed
"go-to bro for all things assistance"
wektor
Well-known member
realtime separation->realtime transfer is definitely on its way, im sure a better tech spec could already handle it
i probably posted this like ten times on this forum but
bear in mind this is two years old now
whats fun is a lot of these work on sort of a compacted representation of audio, ie. you have a net that learns how to compress-decompress audio into maybe 16 numbers or something.
if you have two different sounds encoded into their latent representation, they are essentially points in space you can move between. ie interpolate between think about it and amen break, or subtract them from each other.
will post some demos in a bit for those interested
Clinamenic
Binary & Tweed
as in vector space embeddings?
wektor
Well-known member
afaik autoencoders are a bit deeper than simple vector embeddings but I think the principle of the vector arithmethics is the same, or at least looks like that to me
version
Well-known member
US air force denies running simulation in which AI drone ‘killed’ operator
Denial follows colonel saying drone used ‘highly unexpected strategies to achieve its goal’ in virtual test
www.theguardian.com
Mr. Tea
Let's Talk About Ceps
US air force denies running simulation in which AI drone ‘killed’ operator
Denial follows colonel saying drone used ‘highly unexpected strategies to achieve its goal’ in virtual test
Reports of an AI drone that 'killed' its operator are pure fiction
It has been widely reported that a US Air Force drone went rogue and "killed" its operator in a simulation, sparking fears of an AI revolution - but this simulation never took place. Why are we so quick to believe AI horror stories?
www.newscientist.com