Sign into an existing account

It's gonna be weird when AI is a squirrel that walks up to you in the park and tells you this.

Wait til they can scan things and create realistic replicas. Shit is gonna get weird when you need to ask yourself if youre being followed by a real bird or not.

More like our lives are going to be a bitch because idiots use it to make dumb shit like this. That isn't even believable and wouldn't be surprised if it's illegal.

Until just recent mimicking voices was not a thing we could do. But I imagine we are not far off from that. Giving the voice some emotion will likely require all sorts of specialized settings. Paywalled settings no doubt.

deleted

Yeah. One way is to take some speeches and process them through a neural network that generates the same audio, but with a significant bottle neck in the neural network. That's called an identity transform when you generate the same input and output. When it has that bottleneck in it it's called an auto-encoder. Essentially what you created on the down stream side of the bottleneck is a Hitler voice generator. Now you can likely even keep the same encoder side and it will likely work. If it still sounds too much like you try speaking more like him or loop it through twice.

Now add two levels of complexity to it to make it better. Upgrade it to a variational autoencoder that pushes the bottlenecked layer to be more likely to have a normal distribution among themselves by adding its deviation from a normal distribution to the loss function. Then take it one step further and upgrade it to an untangled variational autoencoder which reduces anomalies. The last one basically clamps the model down to always sound like Hitler no matter what.

https://en.wikipedia.org/wiki/Variational_autoencoder

Now because you are dealing with audio the portions of the neural net to either side of the bottleneck will have to be recurrent neural networks or convolutional because its dealing with time series data. Basically means graphics card on fire lmao.

That would be much more difficult. Different languages use different phonemes. There are some common sounds between languages but without a recording to work from much more synthesizing would be necessary.

The constant cadence and lack of emphasis for emotional effect makes it sound flat. The software will need to improve a lot. AI doesn’t understand what it is reading and so can’t assign an emotion or importance to any particular word. A human will be required to do this manually to the script the AI is reading.

AI can simulate environments .. thats easy

How is this AI?

These are just lines of code and script being read by the machine, coded by humans, not even close to being AI.

Probably not wrong.

You're kinda stupid huh?