Amanda Cupido doesn’t converse Spanish or French, however she and her crew made a podcast in each languages all the identical. Amanda is an audio producer and the founding father of a manufacturing firm referred to as Lead Podcasting. Certainly one of her shoppers is a world nonprofit with over 35,000 workers—and never all of them converse English. So she made them a pitch: what in the event that they added AI into the combination?
Amanda had already been experimenting with generative AI voice instruments when she pitched her shopper on utilizing the tech to make their inside podcast extra accessible to as many workers as doable. They’d make the podcast in English, after which use AI to translate it into Spanish and French—with rather a lot of human oversight, after all. Pushed by a want to make use of these instruments for good, the purpose was by no means to interchange individuals, however to succeed in extra individuals, and it labored.
On season two of the Dropbox podcast Working Smarter, Amanda exhibits what it’s like—and what it sounds like—to make a podcast with AI that’s nonetheless human at its core.
You possibly can learn an excerpt of our dialog with Amanda under.
~ ~ ~
This was your first time pitching generative AI for a podcast. Do you know what you had been stepping into?
No, no. I used to be simply going with my intestine. Generally you say AI and, relying who you pitch it to, they’re very towards it. And I can respect that. However as quickly as I mentioned the phrases, I did not see a grimace. I noticed them lean in, like “Oh, inform me extra.” And so I began simply saying what I knew, which was: There’s know-how that may make this a lot simpler. The workflow goes to be instantaneous. We nonetheless have native language producers on the crew overseeing it and making certain that the standard is as much as par. And the shopper mentioned, “All proper. Get me a value checklist and a proper estimate, and we’ll see.”
So you’ve got pitched this concept to the shopper. You will have this software. You’re able to create an AI-generated model of the podcast in a few completely different languages. How does this course of truly play out?
We’re nonetheless iterating as we converse. You possibly can select one in every of two methods to generate a voice. You possibly can both create a brand new voice profile utilizing the visitor’s audio, or you’ll be able to go to a database of voices and select a pre-established voice that sounds much like the visitor. At first we tried producing new voice profiles utilizing the visitor’s audio, however the translations had been to not the extent we would have liked them to be. What we realized was we would have liked to have a step in-between—so first to tweak the interpretation, and then create the voice. However by episode two, we determined it was truly quicker and higher to make use of the database voices as a substitute of producing a brand new voice profile. There was extra consistency, flexibility, and artistic management utilizing a pre-established voice.
Listening to you describe all the steps concerned, was there ever a degree on this course of the place you or your crew had been banging your heads towards the wall and virtually regretting your determination to pitch this within the first place?
Positively. Yeah, the entire time. The truth is, I did not account for the period of time it was going to take to repair the interpretation. And there have been positively instances once we had been pissed off attempting to construct the brand new voice profiles—therefore why we ended up switching ways. However I gave actual energy to the crew. I simply mentioned, “Go along with your intestine. Do what you assume sounds greatest.” They had been getting nuanced with the accents—is it a Colombian or Venezuelan Spanish accent?—to ensure we had been honoring it as carefully as doable. And so they had been in a position to do this higher with the pre-established ones. So I believed, okay, no-brainer.
It seems like your course of nonetheless required a major quantity of human involvement. Is that right?
For certain. Particularly on the subject of the nuance of translation, language, and accent. That’s deeply private, deeply human, and that human oversight was essential. And in order that’s why the important thing point-people on my crew who had been operating the present earlier than [we adopted AI voices] had been all nonetheless there doing their factor. We simply modified what their job concerned. And as soon as we acquired over the training curve, the workflow was extra seamless. However positively a lot of human fingerprints throughout it.
Why did you are feeling it was essential to take care of that diploma of human involvement within the course of?
From a content material company perspective, it is massively essential to nonetheless have that human oversight. That is what permits us to create glorious content material. And so I do not ever see that going away. We’re all the time going to have that human oversight to make sure we’re getting what we wish from this software. It ended up being so good that once we delivered the primary episode, the shopper mentioned, “Oh, however I believed you had been utilizing AI for these?” She thought it was human. To me, which means it was successful.
A software alone isn’t going to have the instincts or the expertise or all the different information that you’ve got introduced from all the different shoppers that you’ve got labored with up to now. Is that what units you aside from individuals who assume they will simply plug a transcript right into a voice generator and make a podcast?
Precisely. You nailed it. Any particular person can go clickity-clack some buttons and pump out a complete factor and never have that stage of oversight and care. You possibly can’t replicate that.
The place do you land on the utilization of AI extra broadly within the podcasting area?
I actually imagine it is going to assist with globalizing podcasts. That is one of many excellent news tales that is going to return out of all of this. I believe it is going to be a click on of a button on the again finish of your internet hosting website. You select which languages you need to put it out in, after which it is out. And that is nice as a result of there’s so many unbelievable Spanish podcasts that I do not perceive—and vice-versa, English podcasts that others do not perceive. Then we will see, like, what’s the primary podcast? Actually, what’s the primary? As a result of proper now it is skewed. It isn’t correct.
That’s such an enchanting factor to consider.
Proper? After which, from a manufacturing facet, I believe that easy enhancing goes to be made even easier—like, the cleanup edit as we name it within the business. Simply taking out the sneezes, false begins, the rambling. There’s so many individuals who need to make a podcast however do not have entry to the instruments, and that is going to be very easy.
Do you assume that leaning an excessive amount of on one thing like AI voice era takes one thing away from podcasting, or from the individuals who work on this area?
In case you had been to say to the photographer who was in the dead of night room dipping their movie in a bunch of various chemical compounds, “We acquired a digital digicam for you,” they usually’re like, “No, it is ruining the artwork of the darkish room”—we all know how this story ends. We’ll look again on the time the place we needed to do all of this manually, similar to we glance again on actually slicing tape. However that is simply the evolution of know-how, and there will probably be a brand new artwork that emerges that I’d encourage individuals to be open to exploring.
This interview has been edited and condensed. For extra interviews and previous episodes, go to workingsmarter.ai
