00:04:41 · Dan Price, Ezekiel Iferia
Like a couple weeks ago, we did a big presentation to Staples. We probably had eight or nine writers writing on it. They generate 70 scripts. We're looking for one. We end up narrowing it down to the best dozen. We present it to Staples, they pick one, and then that becomes the big campaign moving forward. And that's sort of a good use of us. It's like we have all these great resources that contribute, and then we cast far and wide. Typically in the old days, was New York and LA exclusively. We had recording studios in both of those locations. Now it's all different. Post COVID, nobody does live auditioning anymore, but we orchestrate all of that. And then I'm still the one in the studio working with the actors and the engineer. So, that's kind of what we do. [00:05:41] S0: Alright. Alright. Thank you very much for clarifying that for us. So, know, you saw... Dan, you you saw the rise of AI voice technology, and instead of jumping on the bandwagon, now you noticed a quality gap. So what's wrong with most AI voices right now? You know, why do they sound technically perfect but emotionally flat?
00:06:07 · Dan Price
[00:06:07] S1: Yeah. Another good question. It's a great way of framing it. That's exactly what we noticed, that people were always talking about, Hey, the AI voices are getting better and better. Listen to this and all that. And by people, I mean, yes, some producers and people who are generating social media content and things like that, but mostly the people at those companies. That's where they seem to put most of their energy is they'll introduce 40 new voices, and they think that's their big pivotal kind of contribution or moment is that they keep throwing more and more voices out there. Or they'll say, Hey, we just introduced a sort of lever that you can maneuver and turn this woman into sound that has... She has a French accent, you know, and it's kind of like, okay, that's all great, but how do you use that practically in what we do for a living day in and day out? The answer is you don't. I've never had a spot come across my desk where it says, And mid sentence, this woman starts speaking French. It just isn't a thing. So we had been approached by three of those companies before they launched, asking for our opinion about certain things, whether it was the workflow, how things happened, was the cloning good enough, was the roster of voices good enough, all that kind of thing. Was the price good enough? And in our opinion, it never was any of those things. While the voices are getting more and more impressive that they're replicating or sounding like human voices, for us it always ended there because remember when I said, the way I make my living is in the booth directing actors through honestly, and without exaggeration, it could be 12,000 to 15,000 commercials I've done. So I, of all people, listen to these voices and I'm like, Boy, that copy was not interpreted the way it should have been. Our goal when we cast human beings is the job is always awarded to the person who ended up sounding... They all interpret it well or in different ways, but it always sounded like the person had written the commercial, that they were so connected to it and tethered to it, and it just sort of came out of them so naturally. That is impossible to program a computer to make those judgments and call upon past experience and make it authentic like that. That was the big gap that I noticed ten months ago.
00:09:06 · Dan Price, Ezekiel Iferia
I said, You know, if we can improve upon the voices themselves and make them sound even more human, but bring to our platform performance, the ability to create better performance, that's really going to be worth something because nobody's doing that right now. I never hear anybody advertising, Hey, we provide the tools to get great performances. It's all about, aren't these voices human sounding? That's the easy part. You just put it in the machine and it comes out sounding like a human. But we also felt we could improve upon that because those voices to me lack certain things like this authentic human distinctness that can only be human, at least now. I do think AI is going to get better and better and better, and it was nowhere several years ago. Imagine where it will be in fifteen years. It will sound every bit as much like a human, but it's the performance part that we sort of pride ourselves on providing users the tools to get these performances that we're achieving. So that's kind of how we're approaching the whole thing and our positioning in that space. [00:10:31] S0: Okay. So I like you to, you know, tell us more about Bionic You know, what what exactly... You know, what... What's the difference between what most ear voice sound like, you know, and what BionicVO does differently.
00:10:47 · Dan Price
[00:10:47] S1: Yeah. So it's really just we work harder at establishing the proper baseline for every actor that goes in. So the process is an actor will provide us with just a few minutes of audio. I have a team of four engineers. They've been doing this, working with me for quite a while now, cloning, and have become, in my opinion, the best collection of talented cloning people there are. They end up generating results that actually... Two weeks ago, I heard from a very big talent agent in New York, and she said, That is the best cloning I've ever heard ever, ever, was her quote. And the reason we get those kind of results is we spend on average seven hours working on thirty seconds of audio. We take the three minutes and we're trying to generate thirty seconds of very, very authentic cloned audio so that the actor... We don't stop until the actor says, You know, I played this for my husband, and he said, I cannot believe that wasn't you. Like they're blown away by it. But only at that point then do we start working on the performance. And the performance is very, very tricky. AI, that's why results are not great out there when you don't know exactly what you're doing. We'll clone a person's voice. They sound like me going in and they come out with a British accent. Or there's the slightest pea pop into a mic, and AI thinks you want that in there and it'll sprinkle it throughout. Or it'll listen to a child, think of it as a musical instrument, and start screwing around with the pitch of it. It's very, very tricky. So we study the audio that goes into the algorithms, and if we sense that any of it is throwing off the authenticity or even the performance the littlest bit, we pluck it out here in little pieces, little pieces, until it's just pure audio. And then we continue messing around with the various settings. In the beginning, there's technology where they sort of call it parroting, where as a director, I know how to interpret copy. I know in my head what it's supposed to sound like. I can read copy. I'm not a very good voiceover, but I can interpret. Well, you could feed that in the machine, my audio, and have these voices imitate me. So they have these great voices doing the proper interpretation.
00:13:31 · Dan Price, Ezekiel Iferia
But we stopped doing that because we don't believe necessarily our users are going to have that skill set where they know how to interpret copy in the first place. And so we provided throughout our site, for all I know, it's a 100 tips. They're just everywhere. Try this, make sure you do that, generate enough takes, handle it like a real recording session where you generate eight or 12 takes and take a phrase from here and a phrase from there and edit it together. You can adjust pacing where it needs to be or add little emotive parentheticals, which encourages the AI to say things in a certain way. Write extra words that won't appear in the final product, but get the AI sort of up and running to hit a line in a certain way that, you know, is the way that the copywriter intended. It's a lot of little tricks like that, but that's what we spend time doing so that we establish the baseline for every actor's profile so that when a user ultimately goes in, they're starting from that baseline. They don't have to find the perfect sound on their own. They're they're able to take it from that point forward. [00:14:55] S0: Okay. Alright. Thank you for sharing that with us. So, Dan, you know, let's talk about ethics. You build you build BionicVO around ethical AI practices. So, you know, how is your approach, you know, to AI voice? How is it different from companies that just script voices without permission?
00:15:17 · Dan Price
[00:15:17] S1: Well, we try and be the exact opposite of that. We feel as though, because the very first calls I made when I had this idea were not to agents, they weren't to producers at ad agencies, It wasn't to SAG-AFTRA. All that came later. I started with actors because those... I have friends in the business, dozens of them that I've worked with for five, ten, forty years in a couple cases, and they trusted me. That's why they jumped into this. They didn't even know what the rates were going to be. They just said, You're doing it. I'm in, because I know you're going to handle this ethically, and you're going to handle my cloning with care and all that kind of thing. So we're very motivated to address all of those kinds of issues. What we've kind of built into it are a few things. One is that when an actor comes to be onboarded, they are encouraged to list any sort of conflicts they might have. It could be a specific political party. It could be a brand. It could be a category. Hey, I don't want to do hard alcohol or beer or something like that. It could be anything like that. So as a user goes through and harvests his list of... His shortlist of people he wants to audition, the people who have opted out never even make it to that stage. They're not even hosted as an option to be used in the first place for that. Once the user goes and does some auditioning, that's all watermarked so that it's rendered useless to them. The only time they are given the clean audio to work with is when they've committed to the transaction and actually paid money and hit the button to make the transaction happen, and then in their studio appears the clean audio, and they're then free to generate more takes and import music, and sound effects, and speed things up, and edit, whatever. And we, a few weeks ago, and this is not to say we by any means have installed this, but we've talked to developers about engaging them to create for us an inaudible watermark that would sort of fingerprint all of this audio that leaves our space so that we'd be able to track it, and learn that if someone told us, Hey, this radio commercial is airing in Dayton for three months, and in fact it's airing in Chicago for six on television, that we'll know all about that. So we intend to be very aggressive about those things.