So What? Using Generative AI for Voice Generation

Well, hey everyone. Happy Thursday. Welcome to So What the Marketing Analytics and Insights Live Show. I'm Katie, joined by Chris and John. We're all in one place this week.

Hello. Underwhelming five. Yeah, sorry. This week, we're talking about using generative AI for voice generation. Yes, we are going to show you how to create something that is adjacent to you.

Um, you know, it's one of the one of the goofy games that my husband and I like to play when we're like driving around or you know, watching TV or something, is it AI or not? Because a lot of the voices, a lot of the commercials, a lot of like the if you watch a YouTube video, it's like top eight places to get sushi. You know, a lot of the voiceover narration is AI. Some of it is obvious AI, and some of it is not so obvious AI. So there's definitely some really good tools out there.

You know, a lot of the TikTok and uh Instagram real um no uh voices and sounds. Like there's one that's very clearly like there's a female voice, and as soon as you hear it, you're like, Yep, that's an AI generated, but it they everybody uses it. And so um, we've tried mostly successfully to create uh AI for voice generation for ourselves and we do have a Katie version and so far the feedback has been well that's not you even though it's my voice being used as the training data the output is still not quite me so Chris where would you like to start this week on generative AI for voice generation I think the best place to start of course would be the five P's yay I mean it makes sense it it does and specifically voice generation the technical term is called text to speech so there's two variants of audio there is TTS or text to speech which is turning text into speech and then there's ASR automatic speech recognition which is taking speech into text. The and there's a variety of different reasons why you might want to do this. So let me show you just let's let's talk about what some of those reasons might be.

So if you have content say on your website or things like that and you have people who are uh vision impaired uh you can provide an audio version one thing that people forget about a lot is that it's not just vision-impaired people who find value in audio content. There's also you know call it what you want, call it neurodivergence, call it this or that. There's a lot of people who want to listen to something or they want your content in an audio format so that they can listen while they do something else. Uh, it's why you have things like podcasts and stuff, so that people can say, you know what, I've got a commute, I need to keep my eyes on the road. I still want to read the the Trust Insights newsletter, but I I can't safely read it while I'm driving.

Um, or I'm at the gym, or I'm in the kitchen making dinner, and I and I want to keep up on things, but I I need to be able to listen only. So that accessibility uh matters a great deal. The second set of reasons why you might use text-to-speech is you just need content in all these different formats. Maybe you want a podcast and your podcast uh and you don't have uh a really good microphone setup, or maybe you just don't want your voice on the internet, which is totally fine. Uh taking long form content and turning it into audiobooks.

Anyone who's ever done an audiobook knows how exhausting, exhausting the process is. Uh, 16, 24, 30 hours in a little booth in front of a microphone like this one, just reading your book allowed, and you have to keep up your energy the whole time to have a consistent tone of voice. You know, saying, like, what we don't use GPU cores, Katie. We use GPUs, they all need cooling, right? And and that kind of uh audio book voice.

Boy, does that get tiring? You might have dedicated audio guides, you might want text-to-spee for dedicated audio guides. So imagine you're in travel and tourism, and you might want to generate an audio guide for uh the museum that you operate and and and have people be able to be able to listen, but your museum changes exhibits a lot, and you don't want to have to keep rehiring the same voice out over and over again for like, oh, the bathrooms on the second floor are closed today, right? So text to speech, generate new audio guides, or maybe you would might have uh audio guides in English, and you're like, gosh, when we survey our membership base, 15% of the people who come to our restaurant or museum or whatever, Spanish is their first language. Hey, does anyone on staff speak Spanish?

No, okay, I guess we're doing text-to-speech and translating it. Of course, voiceovers. Uh, and and uh let's uh I'll gotta play a very short uh voiceover um just so that people can give a chance to hear what this sounds like today. So this is from the workshop we just did last week. Your next landmark office or mixed-use project demands more than just design, it demands certainty at Southern Isles AEC, our integrated architecture and engineering teams eliminate surprises, delivering complex, high-performance spaces on schedule, on budget, maximizing your assets' long-term value, experience true partnership and sustainable design that performs.

Southern Isles AEC, building your vision reliably. Learn more at Southern Isles AEC.com. So that's an example of a voiceover and an ad uh that you would use things for. You might have an IVR phone tree, and you might want to upload new files. So if you have a customer service line and you know, you own a restaurant, you might say, you know, thanks for calling our restaurant.

Today's specials are, and you don't want to record that every day because that that gets old after a while. And of course, the final reason is you want to have audio content made available online for AI to train on that it can listen to and and learn from. So those would be some of the purposes of why we would do this. I desperately wish that this technology existed 15 years ago when I was in the throes of creating a computerized version of a substance abuse intake tool because we had a voiceover component to the intake tool. It was computerized assessment.

And so you could shut the audio off, but most people listened to it because it read the questions out loud. And the population that we were serving, the literary literacy levels were lower. So they relied on the voiceover in order to understand what the question was that was being asked. And because it was an intake tool for the substance abuse population, as new opiates and stimulants were hitting the market, we had to update the intake tool with those new drugs and add the same, you know, four or five questions about every single drug. Every time we did that, we had to go back to the agency where we hired an actor to read the question.

And then if the actor got the pronunciation of OPANA or OPANA or however they wanted that specific drug pronounced, we had to go back and have them do it again. And then if that actor decided to leave the voiceover game, we then had to switch the actor and try to find someone who sounded close enough. And so at the end of the day, this product had four or five different versions of a voice, which was very jarring to people because we couldn't get one consistent voice because the actors, the human actors, kept switching jobs or leaving or you know, whatever it was. And so when I see this, I wish, like for me, that was I would have immediately taught myself how to build this. I would have learned how to code just so I wouldn't have to deal with casting and actors and you know, waiting a whole week for something to turn around and having my manager say, why can't we do it faster?

Exactly. Now, I I do think it's important to say, and this is something that we actually said at the workshop last week, uh, from a responsible AI perspective. Um, yes, text to speech is incredibly convenient, it is very, very flexible. However, um we do still want to keep in mind as professionals when we do and don't want to hire an actual human, because we also don't want to eliminate all of the voice actors uh at you know who are out there as professionals. And the general the general consensus is this if you are creating audio content um that is uh a performance of some kind, like an audiobook, for example, and you want to have an ironclad guarantee that you have the copyright on it, use a human.

If you uh because obviously the the individual work itself, you know, the the text presumably is is something that you have rights to, maybe, maybe not. But one of the weird things, and this has come from music, is that the perfor and audio performance is a separate and discrete work from the source itself. That's why you get into music licensing, you have like sheet music, which you have to license, and you have the audio score itself that you have to license and all those things. So from an IP perspective, if you if you want to make absolutely positively sure that any audio you're creating, you have a hundred percent guarantee of copyright on use a human. I would add to that, and this is something that we've talked about, is for the trust insights brand.

I don't want some random AI generated voice being the quote unquote voice of Trust Insights. I want it to be one of us that people know and recognize. And so they can go, oh, that's Katie's voice, or oh, that's John's voice. It must be Trust Insights, because the thing that I see likely happening, similar to how people use the same audio over and over again on social media, is I don't want to pick out a voice from AI and go, that's the voice of Trust Insights, and have our competitor down the street go, and that's our voice too, and have people start to be confused. Now, that's not to say someone couldn't spoof one of our voices, but our voices, we are the voices of our our company, our brands.

So I don't want to let AI be a stand-in for me. Right. So that brings us to the second platform, which is the tools for generation. What are what does the landscape look like? There's three variants.

There's pure text-to-speech tools that are that are dedicated to making text-to-spee. There are hybrid tools. For example, most people uh may not know this. Your operating system, whether it's Mac or Windows, whatever, has a text-to-speech component. You can highlight some text on your computer and say speak it aloud.

And with a very robotic sounding voice, you will get a something that sounds that reads it aloud. And there are tools like Notebook LM, for example, that are not designed as a text-to-speech tool, but they have those capabilities in the same for things like the script, uh, for podcast editing, and so on and so forth. So those that's sort of the the landscape. The within text-to-speech tools, there's three variants. Um, there are tools that are just speech generators.

Google's text-to-speech, for example, is just a speech generator. There are tools that are cloners that have obviously built-in voices, but also you can clone your voice, like 11 labs. And then there are open source tools that offer hybrids of that that have a very very high barrier to entry from a technical perspective to install, but are relatively inexpensive to use. So part of what we're gonna talk about in just a bit is how do you know what to choose? But those are kind of the options for right now.

When you're thinking about what are we going to do or why we're gonna do it, these are the five factors that we need to take into account. Let's see if I can make this bigger because I forgot how. There we go. Um, the five factors are what are you willing to pay, right? If you want you all to pay anything, um, you can use the oh the the one built into your operating system.

Like you can have it read aloud huge chunks of text, and it will, and it will sound terrible. Um, you can also use open source tools if you can figure out how to get them installed. Uh, good luck. I am still wrestling with MLX audio, and I've been trying for like three days now. Uh and then if you're willing to pay some money, uh, and it can be anywhere from you know $25 to $80 a month, possibly more depending on your use cases, you get access to state of the art, but you you get a bill at the end of the month.

For speed, one thing people wonder about is: do you need real time? Or is it a canned product? Like if it's a podcast, it doesn't need to be read aloud real time. If it is an interactive voice guide on your website, it does need to be uh real time. And those there are considerations for that.

How much are you speaking aloud? Any of the web interfaces, the consumer interfaces, they can't handle more than a page of text. And really a lot of them uh have 5,000 character limits and stuff, which is not a lot of text. Uh so if you would need big chunks, like someone asked me what's how do I convert an entire book into an audio book? I said, well, yeah, you you buy uh a subscription to 11 Labs, you upgrade or to Google uh cloud, and you feed your textbook, you feed your book into it programmatically, piece by piece, and you sew together the chunks at the end uh programmatically.

It's a pain in the butt. Uh quality, what level of quality do you want? Uh again, if you're if you're doing trying to do it cheap, it's the old fast, cheap, good, choose two. Uh, what voice do you want? So, to your point, Katie, brand alignment matters.

Should it sound like Chris Panda or Katie Robert? Or is it okay if it sounds like Google chirps uh AOD voice, which is a very nice voice? And the last one, which we alluded to at the beginning, is ethics. Do you have the rights to use those voices? Just because you can doesn't mean you should from a technical perspective.

Uh, I have uh a couple of celebrities cloned in my 11 labs account because I was able to clone them before the restrictions they implemented the restrictions on them, and they did, and they're those voices are grandfathered in. Uh, I do not have the rights to use them. Um, so in any public facing content, I ethically cannot use those voices. I can make stuff for my own amusement, but that's about it. Um, John, I don't know if you've done this, but when I'm writing something like in a word processor, like a Google Doc or a um my gosh, Microsoft Office, they have the ability to you for you to have it read back to you out loud.

So it sort of like falls under the free bad quality. But if you're editing, it's helpful to have something read back out loud to you so you can tell if it sounds coherent or not. Is that a use case that you've ever uh used? Yeah, well, it's just what Chris had mentioned before, recording an audiobook. You know, I've had the manuscripts sent off to two different editors and all cleaned up.

And then still, as I'm doing the audio book, I'm like, oh, wait a minute, that's not right. This isn't done. You know, and so yeah, until you get a human voice reading the thing out loud, um, you're still not done. And uh and yeah, the opportunities for this are just amazing as far as automated stuff, you know. Like you're talking about, you know, weather reports or menus or just anything that is changing all the time.

This could just make all that painful headache go away, as far as having you know, somebody having to sit down and name all the foods on the menu or you know, projects like that. Like this can make that stuff go away. Exactly. When I call my vet's office, the uh chief surgeon at the vet is the one who did the menu. So you have to like, you know, thank you for calling, blah, blah, blah.

If you want to do this, press one. God forbid they change that menu, and that poor guy is gonna have to do it again unless they do something like this. I would imagine that his time is better spent treating animals than recording, you know, telephone menu options. One would think so. Um, so that brings us uh to the process itself of getting this up and running.

If you're going to do voice cloning, Eleven Labs has a pretty decent setup. But this this is true of everything, is garbage in, garbage out. If you record some crap on your iPhone and you bring it into the system, it's gonna sound like crap. Um, there's there's no getting around that. So you need high quality samples recorded of a good microphone, or if you don't have access to a good microphone, uh it is not a text-to-speech tool, but it is a an audio improvement tool, and that is what's called Adobe Podcast.

If you go to Adobe Podcast and go into enhanced speech, um you can hear how how effectively this tool can clean up speech and turn it into something that is somewhat studio sounding. Not perfect, but it can take mediocre audio and give it a nice lift into decent audio. It can't take crap audio and turn it to great audio because again, it's garbage in, garbage out still. But it can take if you if you have nothing else but like your phone, uh you can say, okay, I'm gonna record this, and then and as long as you're click close to the to the thing, you'll get a mediocre recording, and then you can put it into Adobe Podcast and get a good one. I have done this when I've been on the road.

So if I'm recording a video, I'll do it on my phone and then have Adobe Podcast clean it up. I think for a lot of people, that's gonna be a really good option because, you know, depending on what you're doing, what the project is, not everyone has access to a good microphone or even the space to have a good microphone. Exactly. Certainly not something that you want to pack in your suitcase if you don't have to. I am that person.

The next is this is a really important one. For text-to-speech. Machines don't read like we do. You have to process your text for speakability. So that means taking out a lot of function uh formatting and stuff because it will read things aloud like asterisk, dash, etc.

Um, one of the things that I have done to dramatically simplify this process for myself is to turn it into a prompt. So in generative AI, let's uh turn this up so we can all see it here. Maybe to 24. Yeah, this is nice long prompt, and we'll put a copy of this in Analytics for Marketers. If you go to trustinsights.ai/slash analytics for marketers, you can get a copy of this prompt.

Uh, but it basically says go through and fix all these things and produce two versions, one for a human narrator and one for an AI narrator. And the reason for that is because Katie, for example, when I produce stuff for you to read for the in-ear inside for the inbox insights newsletter, my writing is not easy to read aloud. Um it's filled with I thank you for saying it. No, it is not. Exactly.

Um in this this example here, let me just bring this up. This is from this past week's uh newsletter. Uh, you can see how it has gone through and for example, it has turned uh numbers into words. Uh it has uh split out things like the DGX H100 pod, which is written, you know, we you normally write the right text, but it it spells out everything so that the machine will read it properly and it won't sound like total garbage. Um for fun, I've tried the original, it does not know.

They don't know what to do. No, I mean, well, it's funny. Uh, you know, and if you're not subscribed to our newsletter, you can go ahead and get it at trustinsights.ai/slash newsletter. What we've started doing, because you know, as we are trying to market ourselves, we're a small company, we want to make sure that our newsletter is available in as many different ways as possible, sort of as Chris was saying at the top of the show. And so we have the written newsletter, which you can subscribe to, you can watch the newsletter on YouTube and you can listen to the newsletter in podcast form.

And it is me reading said newsletter, but every week I have to get through whatever it is that Chris has written for the data diaries, and I swear he makes it more difficult week over week. It's like tongue twisters. And, you know, because Chris, you write about more of the technical side of things. And so as I'm reading it, you know, in the GPUs, and I was like, am I saying NVIDIA correctly? I don't really know.

We're just gonna go with it. And then I got to the end and I think I was sweating. But it was me, the human. Yes, it was human. Um so once your text is processor speakability, uh, you can send it to the TTS model.

However, uh, I strongly recommend you test small snippets first. Um, like any software development, because that's kind of what this is. You want to test first to make sure that the machines are doing what what they're supposed to be doing, and if they're not, um, you want to to take a step back. So for an example, let's take where is my my uh put that in context. Okay.

So let's go to 11 labs. I'm gonna stop sharing so I can share the actual tab because that's the only way to share audio apparently in StreamYard. Uh let's go to the 11 labs here, and we're gonna use Katie 4.0. Um, and we're gonna have it just read aloud uh this sentence here. Let's see if it will successfully do this here.

Uh to put that in context, that's 71,400 megawatts. The average USA home at any given time is consuming 20 to 30 kilowatts, which means that if AI chips in USA data centers are running full tilt, then AI is using the same amount of power as 2.86 million homes. It's not bad. As you said, it's KD adjacent. Um, however, if I put in the speakability version, it tends to be a little bit smoother.

To put that in context, that's 71,400 megawatts. The average USA home at any given time is consuming 20 to 30 kilowatts. This means that if AI chips in USA data centers are running full tilt, then AI is using the same amount of power as 2.86 million homes. So it's it it handles it slightly better, mostly partly because it's also longer sentences. The machine can uh the models tend to parse that a little bit better.

Um so that's the speakability part. It's really important that we get speakability right before you feed the text to the model, otherwise, bad things happen. Should we do the human versions for the third comparison? Uh I mean you can if you want to. You know what's funny?

I'm glad you brought up that sentence. I stumbled over that second sentence or the third sentence. This means that if AI chips in US data centers are running full tilt, like that part of the sentence I struggled with because to me it wasn't correct grammar, but it could be. And so it's it's well, no, but it's interesting to hear the machine read it. It obviously it sounds a little more robotic because it is literally a machine, but it is interesting to hear it read versus how I was trying to read it.

And I think, you know, it is actually even if the machine version doesn't go live, it is good to have that comparison to be like, am I reading this correctly? Is this what the sentence is supposed to sound like? Or, you know, do I need to go back and rewrite things? So I think it is in in some ways it's a good way to QA. It's very similar to what we're saying about having um, you know, your content read aloud for you know grammar and editing.

If you're trying to read it for like a podcast version or an audio book, having something like 11 labs read it is gonna help you try to figure out, you know, am I putting like the emphasis on the right things? Am I pausing at the right places? Does it sound funny when I do that? This is a good way to try to test that out. Exactly.

So that's the 11 Labs version. Now, um, you can also use there are other services. Uh 11 Labs has an API. Uh you can you'll note here, this window only allows you 5,000 characters. So if the average word is say, you know, seven uh characters, that means that you are limited to about 700 words, which for small blog posts is fine, but if you wanted to have say, you know, your letters from the uh corner office read aloud, it's not going to fit in here.

So you'll have to use the API. Um for Google's text to speech, Google's text-to-speech is entirely based in the API. They do not have an interface for it. However, what they do have are absolutely incredibly good sounding voices. That um if you've used Notebook LM, for example, you've heard these voices.

Um and I will bring uh let me bring up what uh first uh yeah, you heard the the Katie version in uh 11 Labs. Here's what the Google version uh sounds like. In this week's data diaries, let's talk about sustainability. One of the questions that keeps coming up over and over again at events, like the two-day workshop I just did for S M P S A E C AI, is how much of a sustainability impact AI has. The reality is we don't know.

Companies that have massive data centers don't publicize just how much energy they use, but we do know how many GPUs have been sold as a proxy for how much energy AI could be consuming. Nvidia holds something like a 98% market share of GPUS in data centers. So even there, there's a bug, right? You can hear that the plural is just GPUs, right? It's GPUS.

Yep. And then it definitely stumbled over the uh the name of the event because of the way that it was um written out. Exactly. So even in cases where you're yeah, you've used a speakability prompt, uh, it will still have issues if you um, and that's why it's important to do testing of stuff. Any paragraph where you think it's gonna have problems, you put that paragraph in first, have it generate, listen back to it, and go, ooh, well, we need to, I know exactly what we need to clean up for those those particular situations.

Um again, one of the most popular tools that people have used if they wanted to just if they want to do a hybrid, is to use Notebook LM and you can load your documents in there, and then you get the ability to trigger a two host conversation that's discussing your documents. And depending on how you configure it, um, you can give it prompts, for example, like you know, yeah, I start by read saying this is brought to you by trustinsights.ai, and then the hosts of it will be able to read aloud uh their overview. So let me go ahead and share that tab uh so that we can hear what it has to say. Welcome to your deep dive. Ever feel like um digging for buried treasure, like in a mountain of marketing advice.

Oh, yeah, I know the feeling. These trust insights newsletters kind of felt like that. Yeah. But instead of gold doubloons, we're after, you know, golden nuggets. Golden nuggets of marketing wisdom.

Yes. And just like any good treasure hunt, you can't just find the raw data. Right. You get those voices are the Google Studio voices. So if you do not like the way they read uh their their audio digest, uh, within studio, you can have a two speaker uh setup and a lot that will allow you to then uh provide them the exact script that you want, and they will read it aloud back and forth as a as a uh a pair.

I mean, all of this, I find it to be really interesting, and I can definitely see the use case for a lot of companies. Um, especially like if you're like an enterprise-sized company and there is no one voice, you know, that stands in as your brand. This is gonna help expedite a lot of things for you. But if you're a small company like ours, and there is an expectation where it's either gonna be a Chris or a Katie voice, I would have trouble putting something like this out into production and saying this is a stand-in for trust insights because it's not us. And so I think that that's where, as you go back to the five Ps, which is where we started, that's where you really need to figure out like, does it need to be one of the humans on our team as the voice represented for the brand?

Or is it okay if it's a generic, you know, AI generated voice from one of these systems that someone else may also be using? But that's okay because we're generating so much that really the purpose is that you know, it needs to be someone needs to be able to access the audio versus it needs to be a brand voice. Now, one of the interesting things you can do is you can do hybrids. You can do hybrid uh implementations of this, where if it's just you, for example, maybe you're a solopreneur, maybe you don't have a conversation partner, you can uh with basic system audio recording tools, have an interactive conversation. So let me pull up Google AI Studio and Chat GPC has this in advanced voice mode in their mobile app, which you can use as well.

I'm gonna choose streaming mode here. I'm gonna choose Gemini Flash. I'm gonna choose and uh for this, let's use Fenrir as our uh as our um reference. And I'm gonna turn on talk. Uh oh, I've you know, I should probably uh sh stop.

Um I forgot to need to share the the tab so that we can hear what it is saying. So let's go to tab stream Google Studio. All right, and start new stream. Yeah, I didn't like Fenrir's voice. I'm gonna go with AOS.

Hey there, how's it going today? It's going well. Thanks for asking. What about you? Uh I'm doing okay.

I was thinking we could have a conversation about uh marketing analytics, specifically Google Analytics 4 and why the interface is such a hot mess. What's your take on it? Marketing analytics and Google Analytics 4. Got it. It definitely has its quirks.

What specific aspects of the interface do you find most frustrating or confusing? The entire damn thing. It was designed by I think monkeys that were on meth. Okay, so it sounds like you find the entire interface to be poorly designed. What specific issues make you say that?

And what alternative would you suggest? I would suggest using an alternative product like Matomo. Okay, so you're suggesting Matomo as a better alternative. Right. So obviously, this is completely unprompted.

You can, for example, uh use knowledge blocks, which we cover in our Gender VAI use cases course, and drop it in as preparatory material for it to read before the conversation starts. And then have a virtual conversation with this tool. Record it, and now it's your conversation partner. And you might even give it uh an interesting personality so that when it's having a conversation with you, you you can you can be the star of the show, right? You can be like, okay, this is just gonna this is just gonna be an interviewer, it's gonna interview me and make me look like uh amazing.

Yeah, otherwise it's gonna sound like a uh psych student fresh out of college trying to be a therapist. So how does that make you feel? Exactly. Like even just this short 30 second second interaction made me want to throw my computer. Um John, I have a question for you.

So obviously, you know, you are the host of Marketing Over Coffee, and one of the things that I know from recording an episode with you is you record it in chunks. So you do like, you know, the first act, you do an ad for one of your sponsors, because um, if you didn't know, you can sponsor marketing over coffee. And then you do the second act. And I'm wondering for the sponsors, would you ever create a John Wall voice? Because people still have the expectation that it's you.

You have a very distinct voice. Um or you know, my dog, you have a very distinct voice. Would you consider creating a John Wall AI generated voice to read the sponsorship ads so that you could not have to break up the interview into multiple parts? Yeah, there's some sponsors that want it, you know, dead perfect to the script, and so that could work. Um one of the things that a lot, well, of the sponsors that come in, there's a lot of them that want it to be tied a little bit more to the content.

So every week it gets tweaked a little bit as far as what else we're talking about, or if there's anything that that links to it. Um you know, a more interesting one for us because we're such a narrow focus, it's single sponsor, but for a larger podcast, I could see this would be a great way to do dynamic ads. You know, you could have one show, but it's actually serving up a slate of five or seven different ads depending on what the listener's into. You know, you could like there's no point to advertise feminine products if you know the listener's male. You know, you can target that.

So now the advertiser is only paying for the demographic that's going to buy their thing. And that could give you a lot of ways to go. But for us, yeah, you know, a big part of the content is our take on the product and how it fits into the day and things like that. You know, they're not just looking to get the right people and throw them the standardized message, they're looking for a little bit more of a take. Um, but yeah, I, you know, again, I can see how this is, and unfortunately, it's bad news for voiceover artists if you know they're able to do tons of different ads and variants without having to sit through and record and grind through them all.

But you know, I'd argue that that's work that most people don't want to do at all, you know, doing uh the the 65th flavor of a Coca-Cola ad or something is just a tough grind. Well, when can we expect the totally AI voice-generated marketing over coffee episode of Chris and John? I know. Well, we could just run that live one as is there and just keep asking us, well, what do you think? What do you think?

I think that um you could uh you could actually train it diarized on our previous like 10 shows. Train the voice and the speaking style for each, and then use Google's TC, I can then use 11 labs to just ping-pong back and forth recording the segments. Because one of the people things people forget about MP3 files. MP3 files are what's called a headerless, which means you can chop them into pieces and each piece will play as though it was a complete MP3. There's no header or footer.

Like if you do that with an MP4 movie file, neither piece will play. It'll just like, hey, this is an invalid file. MP3s don't have that problem. And so you can cut your text and then, like John was saying, insert a new chunk of MP3 and glue it all back together. Way we're getting like old podcast history, way, way, way back in the day, 2007, 2008, when marketing over coffee first started, there's this company called Pod Show.

And Pod Show, their big innovation was this thing they called Dgap, dynamically generated ad placement. All they did, the hosts had to hit certain timestamps, which I remember folks like CC Chapman are like, this is ridiculous. I have to have exactly six minute segments. But the system was designed so that would drop in an ad made at that moment when someone requested the MP3 file for that show. So as advertisers changed, if you pulled the same episode of marketing over coffee two months apart, you would get a different ad in that file based on the time it was generated.

So you absolutely could do that to John's uh uh point if you had you know that level of detail about your listeners. I think we need the you know, totally AI Chris, AI John episode of marketing over coffee. But I guess, and I don't want to get too off topic because this might be a different conversation. Let's say in this in a world where there's an AI Chris and an AI John just talking to each other, how do you ensure that the information is coming out correct? That it's not just hallucinating off the rails, off the dock, down into the ocean, and then off the planet.

It's the same as every other form of hallucination prevention. The more data you give, the better it's gonna perform, right? Uh so if you if you know, for example, that you're gonna be talking about a topic, say like lead scoring, you'd better provide a knowledge block that's up to date about what lead scoring, how lead scoring works. And that's the thing about all the apps, including the mobile apps. You can provide the data up front before you have the conversation.

You can literally just copy and paste it in from your prompt engineering library and have that conversation, or you can follow the Trust Insights repel framework, which you can get at TrustInsights.aslash repel and have that initial boot up conversation to warm up the device. So I've got OpenAI's uh chat GPT app here. Let's go ahead and turn on advanced voice mode. And let's get near the microphone here. There we go.

How are we doing this morning? Hey, good morning. I'm doing great. Thanks for asking. How about you?

Getting ready to tackle the day? Yes, that's fine. Um, you're gonna play the role of a marketing automation expert skilled at B2B lead scoring. I want you to very, very concisely explain what modern lead scoring practices are great in 2025. Exclude information and sources that are 2024 or older.

Sure thing. In 2025, modern lead scoring leverages AI and machine learning to analyze behavioral data, fermographics, and engagement patterns in real time. It focuses on dynamic scoring models that update continuously, integrating intent data from multiple channels, predictive analytics, and so you can see after you'd follow the repel process, you could turn on your recorder and using in using in this case, speech to speech, have an interactive conversation with a generative AI tool. This is again it's really useful if you are a solo business person, if you are a solo uh podcaster, and you occasionally want someone else to talk to and no one else is available, fire up a text-to-speech app or an interactive AI app and have that conversation. And in doing so, you will create interesting content that people might actually be curious to see.

Like, how do you, Katie, think when you talk to a machine? How is how do you bounce ideas off the machine and brainstorm new stuff? I don't record them because it'd be kind of noisy, but when I go on long drives, like when I travel to the marketing AI conference or to the event I'm going to next week, um, I will do this for hours at a time in the car because at the end of the the the event of the conversation, all of the conversation is logged as text in your AI tool. And then you can just say summarize this conversation in outline format. And maybe you had like 10 ideas for for your next content series, your next podcast.

It summarizes them, it and it it gives you work almost finished work product just from that text-to-spee transcription. But that's not voice generation. That's you yelling at your phone. That is. Well, it's voice generation on its side, not your side.

Right. And so I think what a lot of people, so I think that that's a useful process. And I know that you've written it about it recently on LinkedIn and a few other people have, but I think where a lot of people are stuck is feeling like, oh, I have to use generative AI to create my voice doppelganger. And you know, I think again, it goes back to the five P's of do you really? Do you need that?

Is that a is that something you need to spend your time building right now? You know, we played around with it. We tried to create the Katie version, and we realized that it was just more efficient and effective if IRL Katie actually did the reading because it only takes about 10 minutes, and you're not gonna get, you know, weird things. Um, I tried to have it read something that we created, and it was like, so what? And you can't go back and give it direction and say, okay, I want you to try it again, but I want you to try it a little bit more, you know, whatever.

Like it doesn't work that way. The amount of work that you have to go, you have to give to these machines to get it exactly the way you want it. That's where you have to make the choice of do you need a human or is it machine? And that goes back to those initial use cases, Chris, that you gave uh in the mind map at the top of the show of here's why you might do this. Is it a performance?

Is it accessibility? Is it because your menu changes every day? Is it because you're trying to get something out the door really fast? Are you a small shop? Are you a large shop?

Does it have to go with your brand? Those are all things that you want to take into account because all of this, all of these things, AI, it's all software development. Software development takes time. Yep. So you have to have, you know, you can play around with stuff, but in order to get it out into production, you should probably have a good plan.

There's one more aspect, which is not text to speech, it is speech to speech. Um, and this is definitely getting into slippery slope territory, but we can give this a try because in a tool like 11 Labs, there is what's called voice changer. So I'm gonna go ahead and hit record here. I'm going to read out this text. Enable microphone.

Yes, allow. In this week's data diaries, let's talk about sustainability. Oh, okay. I have to start. In this week's data diaries, let's talk about sustainability.

One of the questions that keeps coming up over and over again at events, like the two-day workshop I just did for SMPS's AEC.ai, is how much of a sustainability impact AI has. And now it will take my original human-led voice and it will try to apply the Katie voice. In this week's data diaries, let's talk about sustainability. One of the questions that keeps coming up over and over again at events, like the two-day workshop I just did for SMPS's AEC.ai, is how much of a sustainability impact AI has. Now, obviously, that is written, that is read with my exact cadence and speech style, but tonally shifted to Katie's voice.

Yeah, that Katie needs to lay off the espresso. Yeah, I've overcaffeinated. That was my first thought, too. Yeah. Exactly.

Well, that's the way I speak. But it allows you to capture more human nuance in terms of how you speak compared to the way the machine interprets it. So if there's something where you have you can match the person's cadence and speaking style, it can apply the tonality to make it sound more like that person. That is clearly obviously still Chris wearing a Katie hat, right? There's no doubt about that.

And I say this is a slippery slope because obviously, again, going back to ethics, do you have the rights to use that voice, right? You can clearly see how this could be very, very badly misused. Oh, yeah. No, I would drive to your house and punch you in the face if you did that to me. Um, however, if we think about in terms of basic use cases, suppose you want to go on vacation for two weeks, right?

And you're like, I'm out of here, peace out. And you're like, well, uh, we could try to front load, you know, three weeks of content all at once, or just say, hey, that week, yeah, AI Katie's gonna read aloud the newsletter, or maybe Chris will read it aloud and and you know, whichever sounds, you know, closer to the actual Katie. I that day I might have to have DCAF. Yeah, well, and it is it is funny because you know, when you think about the distinct brand voices, you and I, we do speak differently, but it is distinctive. And someone could very easily go, oh yeah, that's not, you know, that's not Katie.

Katie doesn't speak that fast at all. Or Chris doesn't speak that, well, I hate to say that I speak slow, but I speak slower than you comparatively. So it's not that I'm a slow talker, but compared to you, I am. Yes. In fact, in tools that are uh ASR tools, like we use Fireflies, one of the things it shows you is uh the number of words per minute.

You speak about 140, and I speak at 192. That does not surprise me at all. Well, and it doesn't, you know, these AI tools aren't necessarily going to be able to factor in. You know, I often say, and I was thinking about this this morning, I often say that I'm a slow thinker and I'm slow to make decisions. And that's not to say that I can't make decisions, but you know, when we're talking on the podcast, for example, and you ask me a question, I often pause for a second because I want to really process the information and think about it and call up any relevant data that I have about it before answering the question.

That's how I speak. That's not how you speak. You call up data faster than I do. And you know, it I don't see how that necessarily translates into you know this AI generation, because that's a hard thing to mirror, because that's you know, very personal to me and very personal to you. Exactly.

And so uh as we wrap up here, one of the things to think about, and I'll have to put this in the mind map, is uh disclosure. Um, you do want to disclose um use of AI because yes, people will know, but if you just close up front and say, hey, Katie's on vacation this week living her best life in uh in the Adirondacks. So you're gonna get the AI version because she deserves to go on vacation. Um, and that will help people understand. Yeah, they it take it removes the doubt.

Is that Katie? Is that not no? It's very clearly machine generated. So wrapping up, this is what we covered today. The main reasons why you would use uh text-to-spee voice generation with AI, the three classes of text-to-speech, right?

There's humans, which is your fallback, hybrid tools, and then pure text-to-speech tools, the five factors for consideration cost, speed, volume, quality, and uh ethics, and then the process for doing this. The prompt, like we said, will be put in the analytics from marketers slack group for the speakability part, and you're welcome to refine that. Um, but any final parting thoughts. I'm gonna start slipping in weird things in my speech so you know it's me and not AI. I'm gonna start using the Matthew McConaughey voice just for everything, just like Instagram.

About you, Chris. Um if people are interested, I might record, you know, because I've got a lot of driving timing up. I'm driving to Toronto for the SMPS uh event there. Uh I'm gonna be driving out to Cleveland for the marketing AI conference. Uh I might start recording some of those AI-based conversations just to see A, how they go, and B, if people would find that valuable, or if people are like, no, I I really don't need to hear somebody, you know, yelling in a car for 45 minutes because in order for the tool to I'm an N of one.

Ignore me. Um or the other option is we could just open up uh our our our mutual friend Chris Brogan used to do this when he was going along drives. He'd just put his phone number on Twitter and just let people call him randomly and stuff. Please don't sign me up for that. No, John doesn't want that either.

That was the old days. But that's gonna do it for this week's episode. Thanks for tuning again. We will talk to you all on the next one. Thanks for watching today.

Be sure to subscribe to our show wherever you're watching it. For more resources and to learn more, check out the Trust Insights podcast at Trust Insights.ai slash TI podcast. At our weekly email newsletter at TrustInsights.ai slash newsletter. Got questions about what you saw in today's episode? Join our free analytics for markers Slack group at trustinsights.ai slash analytics for marketers.

See you next time.