Issue 12 - Deepfake This
Welcome. The Waveform is a newsletter from Red Cup Agency about the podcasts we produce. When we launch a new show or post an episode that stands out, I’ll drop you a note. I’m Lee Schneider, founder and lead producer. Were you forwarded this email? You can subscribe. I’ll write an issue of The Waveform only when there’s news.
Deepfaking a Podcast
A deepfake can take different forms. It can be a video of a celebrity’s head talking while on someone else’s body. It can be Barack Obama or the Queen of England talking trash in a video that is not real, but both looks and sounds real. It can be Mark Zuckerberg explaining how he will take over the world, telling the truth for once — only his voice is speaking words that he never said. That’s the fake part of a deepfake. It’s like Photoshop for videos and audios.
Making a deepfake seemed dangerous. So I thought I’d give it a try.
An app I use for podcast editing called Descript introduced a new feature I could use to make a deepfake of a podcast. Using a relatively limited audio sample of a voice, you could make a clone of it and make the voice say anything you wanted. It was pretty beta, and at first seemed slow and sketchy. Then, as the developers improved it, the voice clone app got really good.
Typing Out a Voice
If you’ve ever written a speech, a podcast, or a one-act play, you know you need to type it out and then speak it, perform it, or record it. With Descript’s audio clone application, called Overdub, you type what you want your voice to say, press play, and the words are spoken in your voice. It mimics the way you speak. It gets the upticks at the ends of your sentences and the emphasis on certain words that you always emphasize. It sounds uncannily like you.
Such wizardry doesn’t come without preparation. You have to train the AI, reading a script to it that it analyzes, or feed it recordings of you talking.
If this all sounds impossibly strange, you could make it become real by having a listen to the latest episode of Same Same but Tech. The entire episode is performed by an AI who sounds a lot like our real host, Máuhan. It captures his cadence and emphasis. The episode, as you’ve guessed by now, is about deepfakes. It includes deepfake experts from MIT and also the inventor of the voice clone app I used to make the episode.
Listen on Spotify, Apple Podcasts, or wherever you get your podcasts.
See you next time on The Waveform.
Lee
Red Cup Agency. Podcast Production.
Working with teams large and small, I take podcasts from the glimmer of an idea into production and distribution.
Read Freely.
You can read freely at The Waveform. I’m not tracking opens, clicks, or forwards. I’m not analyzing your IP address location. It’s just you and me, writing and reading. This is a small part of the vast Web focused on people and dialog, not marketing data collection.