Actor Stephen Fry says his voice was stolen from the Harry Potter audiobooks and replicated by AI—and warns this is just the beginning::The actor told an audience in London that AI was a “burning issue” for actors on strike.
I’m sorry, while I understand this issue is a more visible issue for actors/voice actors, there are a lot of people who are going to be hurt by this in the long run.
You think scam calls are bad now? Imagine if Gamgam gets a call from “you” saying you’re hurt and scared and need money to be safe. And I don’t mean just someone pretending to be you, I mean that the person on the other end of the phone sounds exactly like you, up to and including the pauses in your voice, the words chosen to say, and even the way you roll your r’s. All because someone skimmed your public Facebook videos.
Someone wants that promotion you’re going to get? Record your voice a few times, then have you “drunk call” your boss hitting on them, and then harassing them when they don’t react well to it.
This is going to be one hell of a ride.
It would effectively kill traditional phone usage pretty much overnight as enough people get scammed and scared off the technology.
I’m sorry, while I understand this issue is a more visible issue for actors/voice actors, there are a lot of people who are going to be hurt by this in the long run.
I’m sorry, but as somebody who’s tried out the tech, the amount of vocal processing required is still many hours of data. Even the more professional AI cloning web sites that allow you to clone your own voice require that you submit “a couple of hours” of your voice data. The reason why musicians and voice actors get into the middle of this is because they already have many hours of voice work just out there. And in many cases, the speech-to-text transcription, which is required to train a voice model, is already available. For example, an audio book.
You think scam calls are bad now?
You think scam call centers are going to spend the time to look for voice clips, parse them out, transpose them into text, put them in a model, train that model for many hours, realize the Python code needs some goddamn dependency that will take many more to debug, fix parameter settings, and then get a subpar voice model that couldn’t fool anybody because they don’t have enough voice clips.
They can’t even be bothered to look up public information about the caller they are making the call to. Fuck, the last call I got was from a “support center for your service”, and when I asked “which service?”, they immediately hung up. They do not give a fuck about trying to come prepared with your personal details. They want the easiest mark possible that doesn’t ask questions and can get scammed without even knowing their name.
Imagine if Gamgam gets a call
Who’s Gamgam?
Record your voice a few times
Yeah, sorry, you need more than a “few times” or a “few voice clips”.
Imagine making this post refuting new Ai tech, but being unable to figure out that Gamgam is grandmother.
Shit like this has already happened.
I dunno… have you tried Googling it? I figured it was some word for grandmother, but I’ve never heard of it, and neither has Google, apparently.
Huh, duckduckgo came up with favorite Southern grandma names and 50 best grandma names as the second and third articles. You do have to know what to type in and not always look at the first thing that comes up. I searched “who is a gamgam” and found tons of stuff about grandmas.
Amazon showed off voice cloning over a year ago, and iirc it was claimed to not require hours of content. You’re lagging in your understanding of current capabilities, nevermind the fact that I was talking about the near future.
I have personally used VALL-E and tried it out. What they are claiming is absolute bullshit. It is “a” voice, but it’s certainly nowhere close to “your” voice. Don’t believe me? You can try it out yourself.
Real AI training requires putting in the work.
Well that’s weird, since the article says:
Perhaps owing to VALL-E’s ability to potentially fuel mischief and deception, Microsoft has not provided VALL-E code for others to experiment with, so we could not test VALL-E’s capabilities.
You think maybe that Github release, which isn’t from Microsoft, might not be the same thing despite the name?
All of the AI tech is based on research papers, which are semi-publicly available.
Don’t judge me too much but I have to admit I really like the Warhammer 40k Attenborough channel and there is no way that dude is allowed his likeness
https://www.youtube.com/@AttenboroughLore for the curious and yes, it’s fantastic.
I was watching “him” talk about Tyranids when my kid said, “Isn’t that the voice of the guy who does the animal documentaries?”.
I watched something similar except they did an entire Top Gear episode about Dark Souls 2.
deleted by creator
well lets be real there is a undefinable unique quality to true Original work that most people somehow can pick up on. i dont think ai will ever trully be able… idk whatvim talking about anymore sigh
I’m sure it wasn’t just the HP audiobooks. He’s been on television for 40 some odd years. There’s hundreds and hundreds of hours of recordings of his voice to train an AI model on.
Hours of monologue with zero background noise is absolute gold for training the model though. You’d have to chop up and edit a lot of footage to get an inferior result with the television footage. Still, it’s entirely possible and it may possibly have been trained on both.
I mean humans can do and have been doing this exact thing forever. Computers make it faster and easier, just like everything else. This isn’t AI, this is training a speech model using machine learning techniques.
A few humans imitating other humans is not even comparable to the scale that computers imitating humans can reach though.
True. I think the main difference is that a computer has no moral compass and won’t remember the large scale criminal operation it was a part of. I don’t think it’s worthwhile to fear or regulate this kind of ml application, the cat is out of the bag and the best we can do is implement security controls like passwords with our important relationships.
It needs to be regulated though. The potential to mislead is immense.
This isn’t AI, this is training a speech model using machine learning techniques.
… That’s AI. AI is a subset of ML. Machines Learn to gain Artificial Intelligence.
Maybe, I think of AI as requiring intelligence rather than being controlled by an operator, translating a human voice or text from one style to another. Maybe I’m wrong, it’s just a name after all.
Well, if you think about it, being able to morph an input text into the voice of someone is the job that some voice actors do, and the task of doing such a thing is something that I would define as intelligent.
In any case, it’s not you or me who define the AI term, in CompSci AI is used to mention recommendation systems (which include search engines that do smart ranking, which nowadays is all of them), translation systems, NLP (chatbots, spam filters), Image processing (auto labelling stuff, object tracking, …) and so forth.
Most of these systems weren’t considered “ai powered” before neural networks and deep learning (high layered neural networks), but nowadays most of those are: NLP uses NNs that convert words / text to vectors to then operate around those and decide on stuff (this includes search and translation), Image processing uses NNs too to be able to “learn” information about several images, to then be able to recognize humans, faces, track objects, generate fake images from an input, deepfakes…
Usually, they are called intelligent systems and thus AI because of the way they are prepared, all of those require that some sort of NN is trained with some dataset to then generate a model with weights that is then fed to the NN alongside some input to archieve the desired result. That trained model, which is a bunch of numbers that decide how the input is transofrmed according to the NN, is the, so to say, “intelligence” that the machine as artificially learnt.
That’s it, that’s literally it, disregard marketing and articles written by salespeople, AI is a system that uses some sort of model trained with a dataset, Usually a NN model although in my opinion, any model (random forests for example) should work too, but it’s true that NN models process data of higher complexity.
replicated by AI
Where? Who was selling what?
Honestly, while these growing pains are unfair, I just have no sympathy for rich actors in this. I’d love to get rid of the ridiculous overpaid class of celebrities and operate with artificial stars. I do not enjoy celebrity culture or the elevation of voices and opinions purely on the basis of connections and/or acting talent.
That’s fine. Totally agree, but this (AI actors/voices etc) would just replace the celebrities with just a smaller, richer, more controlling group of capitalist fucks.
Where’s the line for you? Stephen Fry is far from your typical “spoiled rich celebrity actor”. He’s relatively famous, very successful, but like, what dollar threshold destroys your sympathy for people having their careers undermined?
