but more structured
vetting
Monday, December 6, 2021
Well AI Dungeon is the idea, although they trained that on a LOT of unrelated datasets.
link for the lazy: https://play.aidungeon.io
very lewd datasets
Anyways training it on speech first, then developing it for a transcript of urealms live shows I would be able to get a much more refined AI with a few less neurons required
I think its a neat idea. Finally at my pc so I can properly talk
ayy
Hold on let me vet you to the server
training it on speech seems SUPER ambitious
do you have access to some good libraries for language processing?
Anyways I'm currently working on my portfolio, training an AI on a non standard language attempting to produce similar results to Replika with far less required neurons and better memory. So obviously if this works out, then i could apply this to all of the urealms campaigns and see what comes out from it.
You a cs major I take it?
Replying to Spooky Phijkchu #spooky.131 do you have access to some good libraries for language processing?
since i'm working on my portfolio, i need to express that i fully understand the logic. so everything i'm using is written by me. no libraries needed. Also helps massively with optimizations for training.
Replying to Phijkchu_Pikachu #klombak You a cs major I take it?
learned it all by my self
Thats neat
very
talk about projects with potential
NEAT is indeed the algorithm im using lol
most of the upper level academics here are engineers, with me being a mathematical physics major. We have a few people who know some cs stuff though
Replying to Phijkchu_Pikachu #klombak most of the upper level academics here are engineers, with me being a mathematical physics major. We have a few people who know some cs stuff though
i'm home
and we have Mini
our corn farmer
yeah we have the other end of the spectrum too
he'll keep us fed after society collapses
some hs dropouts, and farmers
very wide range server
sick
love small communities anyways
So what's the angle you're taking to mature the AI
are you at the point where you're just feeding it datasets now?
Replying to Spooky Phijkchu #spooky.131 So what's the angle you're taking to mature the AI
I've chosen not to use an "auto" encoder, and instead use an assisted auto encoder. So I spent about 8 months crawling the web for a wide veriety of words. then spent another 6 months crawling every synonym and antonym for all of those words, then the syn and ant of all those... etc.
Then used an adjusted 1d boids algorithm to have every word move closer to similar words. repeated this until no word moved from a regular position in the list for 10 iterations straight.
so now i have 122188 words all with a single value. ie:
yogurt, 0.7066295094612451
normally an auto encoder uses 2-5 values per word. If I can compress this down to just one and the encoder works properly then I can cut down the training time and calculation time by Nth degrees
So technically speaking i would have more neurons, although the AI would be much much faster
Afterwards I'll be training it on a few months of messages between 2 replika AIs to givwe it a quick kick start, then training it on real human data between me and my exes since that is likely the largest dataset i have between 2 people talking. afterwards I will have a trained agent which can be applied to many more specific use cases
Replying to Garbage #garbageatgaming Train an AI off of videos only featuring rob talking, i;e some lords episodes and pokemon nuzlockes to make an AI voice that kinda sounds like rob KappaHD /s
I actually do intend to work on a voice module for the AI, but it wouldnt be truly synthetic, more like AI assisted smoothing of the phonetic audio clips.
we should talk rob into recording every phonetic sound
It's not hard to extract and average phonetic from audio
basically just take the single phoname, find it 100 times. stretch them into place and paste em all over each other. it'll sound like shit but the AI will smooth it properly (hopefully)
It's the same thing siri and alexa does
Replying to Garbage #garbageatgaming So like mixing splicing of old but with much more samples and assisted by an AI to merge them together to make it sound natural?
yup
I know none of this stuff. I know a bit of python from a math and physics background, but I know very little about code and ai
convert the audio into a spectrogram, then have the ai do it's own smoothing filter. to train it you just have another ai which knows humans from robots, when it guesses the robot is a human 50% of the time then you have a good enough voice
exactly
or virtual deadbones
if you're feeling frisky for a weird campaign.
really the only problem is the AI will not really remember what the campaign is actually about
so you could start in the den of devils campaign and end up randomly in the silvermines mountain.
I think that would have its own appeal
it would be quite some time from now as I am still working on improving training times and quality from the algorithm. Right now it is very fast, although it could easily be much much faster. for instance the activation function could be a unique one for this sort of data. since it's highly specific to text and not general data