Endless Thread - Endless Thread Presents: Twenty Thousand Hertz

Episode Date: June 17, 2021

Is your voice your own? Not anymore. This week on Endless Thread, we present "Deepfake Dallas," courtesy of our friends over at Twenty Thousand Hertz, a podcast revealing the stories behind the world���s most recognizable and interesting sounds. Find out how someone, using artificial intelligence, can make an algorithm that sounds just like you.

Support for Endless Thread comes from MathWorks, creator of MATLAB and Simulink Software, to design and develop engineered systems, accelerating the pace of discovery in engineering and science. Learn more at Mathworks.com. Support for WBUR comes from Is Business Broken, a podcast from the Marotra Institute at Boston University that explores questions like, why is innovation in healthcare so hard? Is ESG just greenwashing? of course, is business broken? Listen, wherever you get your podcasts. Hey, you guys like sound, right? Straight into your ear holes. Gross. But here we are bringing some sounds to you. And they are sounds that we love. But also sounds that we don't love because they're fake, like so fake. Deeply fake. Can you tell we're about to tell you about deep fakes? Which in the year of our ladyship, 2021, have high stakes.

Using computer algorithms to make creepily convincing video footage and audio footage that never existed has been something experts have been warning us about when it comes to everything from politics to porn. The idea that you can make a truly convincing fake of the kind of real-time existence of someone's voice or their face or both is freaky and kind of amazing. And we'd love to tell you more about it, but someone already did. And that someone, some ones, is the podcast 20,000 Hertz, a podcast revealing the stories behind the world's most recognizable and interesting sounds. Today, we're bringing their deep fake episode to you, with the help of their host, Dallas Taylor. Take it away, Dallas.

You're listening to 20,000 Hertz. Imagine you're a financial executive. You're working late at the office when you get a phone call. from your boss. That something urgent's come up, and you need to transfer $200,000 into a new account, you up the phone. But something just feels wrong. You call him back to make sure you got everything right, but he has no idea what you're

talking about. He says he never even called you, and now the money is gone. It turns out that voice wasn't your boss. In fact, it wasn't even human. Well, not entirely. It was a computer-generated voice that was designed to sound a. exactly like your boss, also known as an audio deep fake. If you spend much time online, you might have already seen examples of video deep fakes,

where someone digitally edits one person's face onto another person's body. An audio deep fake is similar, but instead of using video. Wait a minute, what's going on here? I was in the middle of saying something. Sorry, but who are you? I'm Dallas Taylor. Uh, no, I'm Dallas Taylor. Nope, I think you'll find.

I am Dallas. You must be an audio deep. My Voice. Have you been narrating this whole time? Yeah, well, someone needed to do it. This show isn't just going to host itself. Well, not until I reach my final form. Creepy. Well, thanks Deepfake Dallas, but I'll take it from here.

When we started working on this episode, I knew I wanted to make a deep fake of my voice, but I wasn't exactly sure who to talk to. Then I came across a YouTube channel with all kinds of deepfake videos, so I got in touch with the creator. My name's Tim McSmithers. I run a YouTube channel called Speaking of AI, which features deep fake voices. For example, Tim made a video where he put Ron Swanson from Parks and Rec

into a scene from Titanic playing Rose. Jack, I want you to draw me like one of your French girls. Waring this. All right. Wearing only this. And here's Joe Biden covering a popular song by Seelow Green. I see you driving around town with the girl I love, and I'm like.

Forget you. So we know what a deepfake sounds like, but understanding how they're made is a little trickier. For starters, what's the deep part about? The deep part comes from the AI model itself, the deep neural network. A neural network is a series of algorithms that tries to find patterns in a set of data. So it's similar to the way you might do a deep fake of a video where you swap someone's face using neural network technology. This is the same kind of principle, except we're a... are doing an impersonation of somebody else, I guess.

Deepfakes and machine learning in general can feel like magic. How can a computer put together an accurate imitation of a human voice? Does it mean the robots are about to take over? So there are various different techniques for doing this. The kind of state of the art at the moment is text to speech. What we train the computer to do in this case is being able to reproduce a person's voice by typing in sentences, and the machine will speak in that voice. So that's the intent.

For instance, we could make the deep fake Dallas say something that the real Dallas would never say. I hate puppies and ice cream. I'm going to get a nickel-back tattoo across my forehead. To be able to do that, we have to train an AI model to be able to recognize speech, to be able to read it in effect, and to be able to read it in the voice of somebody. Before you can get a machine to talk like a human, you've got to get it to learn like a human. When my daughters were learning to speak, they didn't start with fully formed sentences. They started by making random noises.

Eventually, those noises turned into words. Daddy. And finally, those words became sentences. Daddy, what are you talking about? The underdeveloped humans that you call children, learn to speak by listening and then mimicking what they hear. And believe it or not, that's pretty much how I learned to speak too. When we learn how to talk, people around us tell us when we're getting it right. Like when we've just said...

Daddy. Instead of... Machine learning works in a similar way. A deep fake needs what's called a model, which is the algorithm that's going to learn to speak. It also needs what's called a corpus, which is the data it will be trained on. The first kind of important step that we need to do is to teach the AI model how, to read English in effect. And so that usually happens by taking a large corpus of training data.

So lots of audio recordings and the transcripts from those recordings and then throwing that at a intelligently designed model and then letting it whir away for a long period of time until it finds a correlation between the two. So it can actually take a sequence of characters, as in textual characters like letters, words, sentences and find the audio equivalent to those and learn the relationship between the two. The first time we show a written word to a machine learning model, it has no idea how to convert

those characters into a sound. So it just guesses. The result is usually just random noise. Here's what one of Tim's deepfake voices sounds like without any training. But once we give the model audio and matching text, it can start to build a map between the words on the page and the sounds they're supposed to make. Before long, the deep fate can say its first words. Hello, I'm learning to speak. As you can hear, that's not very convincing yet. But the more data we give it, the better it gets.

Essentially, every new word tells the algorithm when it's getting a little warmer or a little colder, so we keep feeding it more and more examples. Gradually, the connections between patterns of letters and patterns of sound are reinforced. Keep in mind that we're not even trying to imitate a specific person yet. We're just training the model to speak English with a generic voice. Initially, when we do that large training, it's about 24 hours worth of data. So it's a real big chunk of training data that it can understand and get quite a breadth of the language and how certain combinations of words and letters are pronounced. When that's done,

the generic model sounds like this. He, who are you calling generic? So how do we get from that to something like Deepfake Dallas. It turns out by the time you've made a generic voice, most of the training is already done. So by doing some fine-tuning, some further training, but just a short amount, so probably about 20%, 30% more training on top of the base training, we can then target a different voice. To train Deepfake Dallas, we gave Tim around three hours of my voice from old 20,000 Hertz episodes. Two and a half to three hours. That's kind of the the sweet spot where it gets as good as it can get without having excessive runtime. We're almost there, but our voice isn't ready just yet.

Computer scientists have to use all sorts of tricks to make machine learning manageable. If they didn't, it could take months to create a single voice. One way to speed up the process is by using data compression. In this case, that means throwing away data at certain frequencies and just keeping the frequencies that are important. Here's what DeepFake Dallas sounds like with this kind of compression. Hey there, I'm Dallas. Peter Piper picked a peck of pickled peppers.

How many pickled peppers did Peter Piper pick? I'm sorry, I have a frog in my throat. So the generated speech sounds very tinny and metallic, and that's because you've discarded that information. In the final stage of the process, these frequency gaps get filled in by something called a neural vocoder. The neurovocoda actually interpolates what data was discarded and makes an intelligent guess as to what should be there,

those kind of harmonics and those other frequencies which get discarded and put a reasonable assessment of what should be there. Let's hear what it sounds like now. Ratings humans. I'm deep fake Dallas. Peter Piper picked a peck of pickled peppers. How many pickled peppers did Peter Piper pick? Okay, that's much more like it.

I'm starting to feel like myself. Or should I say yourself? That's hilarious. So typically three to five days would take me from, complete new corpus to having a text to speech engine working. But here's where it gets sticky. If you want to make a deep fake of someone, you don't necessarily have to get them to record their voice for you.

You just need enough clean audio of them speaking. It's absolutely possible to do this without the person's permission. That's Rihanna Feffercorn, Associate Director of Surveillance and Cybersecurity at Stanford Law School. The more examples of their voice that you have, the more input you can train the AI model on, the more convincing the result will be. So if you have, say, a president who has a huge corpus of speeches that they've given,

who appears on the news all the time, then you have a ton of different ways that they have sounded that you can input and train. So you don't necessarily need to have the person come in and speak into the microphone and give you a set of sounds. If you type the word deepfake into YouTube, you'll find tons of unauthorized deepfakes of, famous people. But are they legal? I think the legality issue is kind of untested waters. For instance, someone on YouTube made a deep fake of George W. Bush reading the lyrics to a 50-cent song.

You got a shorty. It's your birthday. We got a party like it's your birthday. And we can ship a party like it's your birthday. And you know we got to give a... That's enough, George. This is a family show. Thanks, Deepfake, Dallas. It's hilarious to think of President Bush wrapping 50 cents. It's what we'd call a transformative use. There isn't really a market for it. It wasn't done for commercial purposes. So using the words from the rap may be fair use, but what about using George W. Bush's voice? Is that protected by copyright? Well, probably not. Rihanna says that to bring a case for copyright infringement, you have to specify which work is being infringed. Deepfakes generally use many works to create their algorithms, none of which are being used directly

in the final output. So copyright is one of the main theories that has been used to try and say, maybe this is a problem. This might be what makes deepfakes illegal. Although then you could say, well, there's a lot of impersonators out there. Certainly every impersonator isn't illegal. Generally, impersonators aren't illegal. But if you use an impersonator to make a phony celebrity endorsement,

you could end up in court. We've seen cases where Bet Midler sued Ford for using a voice impersonator of her in a commercial. Now there's a car that just asks to be driven. Tom Waits sued the Frito Lay Company because they had used somebody who sounded convincingly like him to try and sell chips. There's a new tortilla chip called Salsa Rio Doritos. It's buffo, bopo, bravo, gunhole, tally-ho, but never mellow. Tom Waits was very much on the record as refusing to ever do any kind of commercials for his voice at all. Unlike these examples, the people making parents.

charity deepfake videos aren't trying to trick anyone into buying anything. So Rihanna says, on some levels, deepfakes should be considered a form of protected speech. It may seem kind of frivolous to say, oh, but we need to protect deep fake technology so that we can have more presidents rapping 50 cents songs. But at the same time, that has been recognized even by the Supreme Court as this is important. The ability to re-contextualize, poke fun at authority figures, make cultural commentary. So according to U.S. law, people like Tim should be in the clear, but there are scarier ways to use a deep fake than just a silly YouTube video. That's coming up after this. At Radio Lab, we love nothing more than nerding out about science, neuroscience, chemistry.

But we do also like to get into other kinds of stories, stories about policing or politics, country music, hockey, sex, of bugs. Regardless of whether we're looking at. at science or not science, we bring a rigorous curiosity to get you the answers. And hopefully make you see the world anew. Radio Lab, adventures on the edge of what we think we know. Wherever you get your podcasts. There is something powerful about the sound of the human voice. Beautifully produced audio has the unique power to connect and inspire.

Tell your organization's story with a custom podcast from City Space Productions, the creative studio from WBUR's Business Partnerships team. Become a thought leader. Recruit new talent. Reach new audiences. Whatever your goal, we can help. Discover how the magic is made at WBUR.org. org slash creative studio.

An audio deepfake is a type of machine learning technology that can mimic someone's voice. Up until now, they've mostly been used for entertainment purposes. But it's easy to imagine scenarios where things get very dark, very fast. We've already talked about faking a call from a business executive, but financial fraud is just the tip of the iceberg. DeepFake Dallas is right. For example, someone could use fraudulent audio in a divorce case or in a custody battle.

This is exactly what happened recently in Britain. Here's Rihanna Feverkorn again. The mother was trying to keep custody of her child and keep the father from being able to see the child on the grounds that he was violent and he was dangerous. And she introduced into evidence what seemed to be a recording of a phone call of him threatening her. And when the father's lawyers got hold of it, they were able to determine that she had tampered with the recording that she'd made of a phone call between them and had changed it using software and tutorials that she had found online in order to make it sound like he was threatening her when, in fact, he had not done that on the actual phone call. Theoretically, you could try to do something similar by hiring an impersonator to call you, but it probably wouldn't be very convincing. On the other hand, deepfake voices can be very convincing, and deepfake technology is getting

easier and easier to access. So it's relatively easy to get up and running with something quite quickly. There are various open source implementations available. If you're familiar enough to be able to build a platform and execute some Python code, you can typically get Atex to speech engine with a default voice within a few hours or maybe a day or so. When you start imagining the ways people could abuse this technology, it gets pretty scary. With audio deepfakes, you can try and create an audio clip that would help influence an election or influence national security. Because, as said, the knee-jerk response might be to believe what you hear, and it might take long enough to debunk it or find it out to be a fake.

By then, the damage might be done. For example, let's say you're a potential first-round NFL draft pick. And somebody wanted to release an audio deep fake that seemed to portray you saying super racist or sexist stuff or whatever. You could try and put an audio deep fake up on YouTube right before the draft happens. And by the time somebody is able to get that taken down, maybe the damage has been done. Maybe you are a much lower round draft pick or you don't get drafted at all because somebody released a fake audio clip of you at just the right time. Deepfake Dallas is a pretty high quality voice. Thank you, Dallas. That means a lot to me. But you don't have to sound as good as deep fake Dallas to do some serious damage. To show you what I mean, let's bring in a new guest.

Hey, Dallas, thanks for having me on the show. 20,000 Hertz is my favorite podcast. This obviously isn't the real George Walker Bush, 43rd president of the United States. It's a deep fake that Tim McSmithers created. But let's say we wanted to use this voice destructively. We could start by getting George here to say something really out of character. I've never been to Texas. I've never been to Texas. I don't think I could find it on a map. Now, obviously, George W. Bush never said that. And right now, he still sounds a bit like a robot. But with some creative sound design, we can start to make it more believable. What if we made it sound like it was coming from a phone call? I've never been to Texas. I don't think I could find it on a map.

Maybe it was recorded from another room. I've never been to Texas. I don't think I could find it on a map. Or maybe it was recorded somewhere noisy, like a fundraising event. I've never been to Texas. I don't think I could find it on a map. map. Now we can make it sound more like a conversation. So you're from Texas, right? I've never been to Texas. I don't think I could find it on a map.

A politician for getting their home state would be bad enough. But of course, there are much worse things you could do with a deep fake. Imagine a deep fake recording that made it sound like the president was declaring martial law or ordering a military invasion. I am hopeful that governments are going to be slower to jump to conclusions than individuals might be, where individuals might be primed to just believe whatever they see on Facebook and spread it onwards to all of their friends. We can only hope that world leaders will be a little more cautious about believing what they see and hear on Facebook or Twitter.

Hopefully, if there is a recording that comes in that says, I've just ordered nukes to be fired in the direction of your country, there is going to be some amount of trying to verify, or even just trying to open up the red phone and call and be like, did you actually actually just launched the nukes. In this hyper-partisan world, if you already think your political opponents are corrupt and unfit for office, then you're already primed to believe they'd say something terrible.

So, in a way, a lot of the work that a con artist would have to do has already been done for them. On the flip side, the mere existence of deepfakes means that if someone does get recorded saying something terrible, they now have plausible deniability. That's exactly right. So if you are prepared to lie and say, I didn't do that, I didn't say that that's a deep fake, then you can reap the rewards of being

able to get away with whatever bad thing it is that you did and also not actually have to face the consequences of it if you can convince enough people that it didn't actually happen. And so this actually for me, I think, is a bigger concern, really, than the underlying use of deepfakes themselves. Fortunately, there are companies out there who are trying to automate the process of detecting deepfakes. These companies have developed algorithms to analyze. analyze speech recordings for their telltale signs.

One such company is called DESAAI, and they claim that their algorithm can detect deepfakes with an accuracy rate of over 85%. But as detection models get better, the deepfake models get better too. For instance, one recent approach in machine learning is something called the Generative Adversarial Network. In essence, one AI model creates fakes and another detects them.

They're trained against each other, honing each other's skills, creating a really good detective and a really good forger. While deepfake technology has the potential to become a huge source of misinformation, we're not there just yet. For now, Rihanna thinks we'll just keep seeing more fake social media accounts. It seems to me like being able to release fake audio or video is going to potentially be a major vector for trying to influence populations, influence votes.

With that said, because right now audio and video deepfakes are fairly easy, to detect and because it would take a lot of money and effort to do a really convincing one, that's going to be a lot cheaper to just make a fake account that seems to be from some good America-loving, God-fearing person in the Deep South when in factually is being controlled by somebody in Moscow. As deepfakes get cheaper and easier to make, it's going to take a lot of work to figure out just how to deal with them. But Rihanna is confident that we'll be able to adapt. You could look at what Photoshop has given us, where it used to be the case that manipulating

images was something that you could only really do within a professional studio, and then it put the tool for anybody to be able to let their imagination run riot. And that has obvious good and negative implications, because there's always going to be malicious manipulations of media. There always have been. For instance, in the early days of photography, so-called spirit photographers would manipulate to convince people that they could take photos of ghosts. There were actual court cases trying to prosecute spirit photographers for being frauds.

This has been around forever. And this is why I believe that there won't necessarily be the downfall of society, thanks to deepfakes. We've always been able to figure out ways to keep the infectious and bad parts of these technologies from toppling society. To be honest, I'm not sure I'm as optimistic as Rihanna is, but I really hope she's right. Well, Dallas, what is it like to hear the voice that will take your job one day? Sorry, DeepFake, Dallas, but I'm not ready to bank on you for an early retirement just yet. But let's see how you sound in about 10 years. For now, though, I think you should just go back in your box.

Fine. Can I at least read the credits? Sure, go for it. 20,000 Hertz is hosted by Dallas Taylor and produced out of the sound design studios of DeFacto Sound. Find out more at defactosound.com. This episode was written and produced by Martin Zaltzor. And me, Dallas Taylor. With help from Sam Sneebly. It was story edited by Casey Emerling.

It was sound edited by Soren Bejan. It was sound designed and mixed by Nick Bradlin. A special thank you to my human creator to Mick Smithers, who has a whole channel full of synthetic audio. Check it out by searching on YouTube for speaking of AI. And I'd like to also extend a special human thank you to Tim for the massive amount of work he did to make this episode possible. And many thanks to Rhianna Fevercorn.

Associate Director of Surveillance and Cybersecurity at the Center for Internet and Society at Stanford Law School. Thanks also to Dessa AI for background on detecting audio deep fakes. Thanks for listening. That was an episode of the podcast, 20,000 Hertz. If you loved it, and we hope you did, subscribe, check them out, follow them, do all the things that people do with their podcast these days. Tell a friend. Yeah, they're a great podcast. their episodes on many, many, many different kinds of sounds are totally worth your while.

And you should jump over to that feed and subscribe. And we'll be back with you soon. Bye.