Second in Command: The Chief Behind the Chief with Cameron Herold
00:24:30 · Cameron Herold
[00:24:30] S3: Yeah. I was at the... I've gone to the main five day TED conference for eleven years and at TED, I think it was last year. It was either last year or three years, it must have been last year, they had this guy on stage singing and then all of a sudden his voice changes and it's this woman's voice and you see his mouth moving, you see him singing into a microphone and AI literally changed it to this woman's voice in a different language. We're like, what the fuck? This is like... It's really cool shit. Like this isn't... And and what I'm really frustrated around people speaking about AI right now is they're all talking about AI and they're talking about one software. They're talking about ChatGPT. That's one tool. I I have a dashboard called There's An AI For That and there's something like 8,000 different AI tools that exist and everybody's talking about one of them. So how do we get people to understand what Eleven does and to play with the tool? Who would you want playing with the tool to see the opportunities?
00:25:28 · Victoria Weller
[00:25:28] S1: Yeah. I do think that OpenAI has like a pioneer in creating that fundamental shift or creating that that mind blowing moment for people first interacting with ChatGPT and understanding, like, AI is here to change the world. Yeah. But then I think from the text space where ChatGPT operates to the visual space where DALL E or Midjourney operate, and then the voice space where ElevenLabs dominating at the moment. And I think the people to your question of who should be using it, I think there's also almost no limit to this. The immediate thought is content creators. So for example, you with podcasts or audiobooks, but also with shorter narration of videos or shorts even, like the voiceovers for any length of video or movie or film can be done with ElevenLabs. But then you can continue thinking about it, and if you have awareness just going through your day, you can see almost infinite use cases for text to speech. One of the most unique ones I found within our customer base is a hospital that uses ElevenLabs to read out before a surgery all the possible side effects and consequences and everything that has to be read out to the patient before Wow. The surgery. And for that, also, you understand why you wouldn't want it to be too robotic, and maybe even a doctor that reads this has to, like, rush through it and do robotics. So a really good use case for that, but not one that I would have thought of in a dark room coming up with with use cases. But one that I'm also really excited about is for accessibility. So we found that on the b to c side, there are many users that use ElevenLabs for text to speech, be it because they can... They're visually impaired and and they have a hard time reading. And it's so much nicer to have a human voice reading things to you rather than a robotic one, especially if you spend the majority of your day with text readers, etcetera. Or people that have lost their voice as a result of throat cancer or similar and or have trouble, like, with stutters or finishing their sentences. There's so many physical things that can impede your speech where ElevenLabs unlocks the option to speak again and even to speak again in your original voice. For those that have, like, changed their voice over the course of time, they can have their old voices back or even their their voice back, period.
00:28:08 · Victoria Weller, Cameron Herold
So I think in the accessibility space, it's also super exciting. That's on the b to c side, and we also serve to enterprise customers. And there, it's another... You open a Pandora's box. And when there's just so much in there, like, from automotive to, like, that hospital use case to gaming studios to production companies, everywhere. The first moment I had with ElevenLabs was when I heard about it, and then I took the bus home. In every station, there's a voice that announces the next stop. And my mind just went to the operational costs of creating these recordings for me to know which bus stop is next. And then there's a diversion and suddenly the bus driver has to tell me what the next stop is. And somehow, it's almost impossible to understand them. And and so for me, I was like, okay, transportation, huge use case. Probably a smaller one if you look at our enterprise suite now, but even there, like in in day to day life, there's so much text to speech that could be leveraged. [00:29:17] S3: So I just thought of a bizarre use case. My mom passed away twenty years ago and there's an email that she wrote me that I reread frequently and it would be wonderful to be able to hear it in her voice and that would be something that if we ever had her voice, that could easily if I had videos of her speaking, I could train AI to be able to speak and then have her read the email to me of...
00:29:35 · Victoria Weller
[00:29:35] S1: Yeah. You know what? That is actually a use case that comes up a lot for for mothers, but also other family members. And we just implemented a a workflow specifically for that, for the compliance parts of it. There are two ways on ElevenLabs to clone a voice. The first is instant voice cloning where you upload thirty seconds of audio and then you get a voice clone and you can read any text out loud in this specific voice. And then there's professional voice cloning, which is a higher quality voice clone for which you need thirty minutes of input data. And because we want to ensure that people cloning those voices have the rights to the voice, that it's their voice, we have a voice capture as part of the flow. So you upload your samples, and then there's a flow in which you are confronted with a random sentence and you have to read it out loud. And then we manually go through to ensure that those voices are the same, And only if that proof exists that it's your voice, you can go... We allow the model to train on the uploaded data. And if you wanted to do that for your mother, there would be no way to pass that voice capture. So... And we've implemented a workaround where if you provide sufficient information to prove that you are like the next of kin or have per permissions over your mother's voice, we allow to professionally train, professionally clone those voices as well.
00:31:06 · Cameron Herold
[00:31:06] S3: That's incredible. One of the questions I had written down was around the legality of it and the prevention of the deep fake videos and audio. One of the big scary, I don't think it's a meme, I think it is actually happening, are these audio messages like voice messages that someone is getting from their mom, you know, I'm in trouble, don't call me, don't email me, send money to this number or they're going to kill me, and it's a fake. It's literally Like, how are we preventing those kinds of things from happening? Is that is that from the training?
00:31:41 · Victoria Weller, Cameron Herold
[00:31:41] S1: There's so many aspects of this and definitely something also that's really important to think about and to continue thinking about as AI gets better in all the different facets that exist. And I think a huge component of this is education. Being aware or creating that awareness in in the public that AI can now be almost anywhere and to have a critical eye when being confronted with a a piece of content. The first pieces of this, I think, I remember when Photoshop was being used more and, like, you would see posters of photographs and you would think, like, oh my god, these people are flawless. And then they'll be like, no. They're... It's... Things are photoshopped. And I think the next shift of seeing or hearing content should be to to think like, is this real or is this generated? So I think there's a huge educational component. But, of course, a single company, like, it would be lazy to say the public just has to be aware and then we solve it. There are also safeguards that we can roll out as as part of developing the tech. And I think every AI company that is innovating at the moment has an ethical responsibility to develop safeguards. [00:32:55] S3: Is it like watermarks? Like the kind of watermarks that