How I Actually Make AI Voice Sound Real

[00:10] So, in this vid- no, wait. So, in this video,  we’re going to… break down the… Ugh, that   sounds terrible… Still recording? 

[00:22] I can’t get it right. I just  don’t sound… like a youtuber.  Good? What’s good about that? Have you heard an advice,   “if you’re homeless, just buy a house”. What the f? That doesn’t make any sense. 

[00:37] I mean if you don’t sound like a youtuber,  then just use something that does. And just like that, I went down the  rabbit hole of… AI voice generators.

[01:08] So… this is how I actually  make an AI voice sound… real… Okay, look. Before we get into the technical  stuff, if you wanna make a realistic voice with  

[01:22] AI, you need to understand what the hell makes  human voice realistic in the first place. Really,   like why does your brain think you're listening  to a real person right now instead of a computer   pretending to be a ginger dude with a fancy suit?  Well, if you break it down, there are 4 elements.

[01:37] If I keep the same tone throughout the entire  video with no ups and downs in my voice,   that would make it sound robotic. I  talk like that in real life… Oh, man,   this robot can talk more human than me?  Anyways! If all of a sudden I change my  

[01:50] tone from monotonous to super excited,  that would make it sound more human. You probably heard that cutting all the pauses  in your video boosts retention. Even though  

[02:02] that’s kinda right, it’s not worth making  yourself look like a lifeless robot that’s   been controlled by the algorithms. Instead,  pause on some important points… to make your   message… more impactful… And make your  AI voice sound more natural, of course.

[02:19] Just like making pauses, emphasizing important  words in your script… no… EMPHASIZING important   words in your script, is also a powerful way  to make your AI voice sound more realistic.

[02:32] And lastly, of course, if you use  ChatGPT to write your entire script,   no matter how realistic your AI voice is, at some  point, it will start to sound robotic. And yeah,   if you write your scripts with AI, and  think that nobody knows… trust me… we know…

[02:48] Alright, so with these 4 elements, you  can cook a 96.2% pure ai voice. These   are the main things you need to keep in  mind while creating your voiceovers. And   But before that, we need to…

[03:03] Look, there are so many AI voice generators out  there right now that I just gave up counting. I   tested almost every single one for this  video, and you know what? Most of them   are… fine. They do a decent job at replicating a  human voice. But the one I always come back to,  

[03:19] is called, ElevenLabs. So that’s gonna  be the software we’ll use for this video. And no, they didn’t pay me to say  this. I wish they did, but they didn’t. and you’re seeing… or hearing how  realistic it sounds. Be honest. You  

[03:34] sometimes forget it’s not a human voice,  right? You… you don’t? Okay, nevermind… check it out and follow along with your  own script. Just create and account,   and try the techniques we’re gonna use  by yourself. Okay, fine. Let’s do this.

[03:49] So you know the theory. Now,  what do you need? A voice! This is an important step. Because once you  pick a voice, you're basically married to   it. All your future videos. That voice. So  don't just click the first one you see and  

[04:01] call it a day. You’re gonna need something  solid. Take your time. Pick a nice one. First, if you wanna use your own  voice, you can go with this route.   You may think “why would I clone my own  voice when I can just… use my throat”,  

[04:17] but it’s gonna be a really important step  for the end of the video. Just let me cook.   Click the ‘voices’ tab. ‘clone a voice’, then  ‘instant voice clone’, and upload a sample.

[04:29] But keep in mind that it requires at  least 30 minutes of voice sample. Once   you generated your voice clone, “you can  use it wherever you want. that’s crazy”. Wait, actually… yeah, that’s completely illegal.  

[04:44] against every terms of service ever written.  So you better use it with your own voice only. And, uh, Finzar, if you’re watching  this… please don’t sue me, dude.

[04:56] seems like no ai can say it like the original you can choose one of the voices  from the huge voice library here.

[05:20] after “how you edit your videos”, is “which  voice are you using in ElevenLabs”. Well,   the voice you’re hearing right now, isn’t  from the voice library or a cloned voice…

[05:32] It’s a custom voice! Most tutorials don’t mention  this, but you can actually create an entirely new   unique voice in ElevenLabs. It’s usually as  realistic as the ones from the voice library,  

[05:44] while being not available to everyone with  one click. So it’s a great choice for creating   a brand “that doesn’t use this voice”. No  one else can use it for their channel. Uhm,   well, if they don’t clone your voice with the  cloning feature, but you get my point, right?

[05:57] In the ‘voices’ tab, click ‘create a voice’,  ‘voice design’. When I created my voice a few   years ago, this feature was just a single button  that you can click, and hope for a good voice.   But now, you can describe the voice in your mind  however you want. Like “a young man in his 20s,  

[06:12] and variances in his speech, changes the tone  and speed continuously throughout the speech”. I   don’t know, something like that. And let’s reduce  the guidance a bit to see what it comes up with.

[06:33] Damn, can I change my voice to that? No? Okay… Now, if you found a voice that  works for you, we can move on to… At this point, you have a nice voice, and you know  what realistic voice sounds like. So you can just  

[06:48] copy your script, paste it to Elevenlabs, and hit  ‘generate’ wait, hold on for a second. Of course,   you can do it. You may even get a pretty good  result from it as well. But if you do that,   after listening to it for a few minutes,  something will feel off. And the viewers  

[07:03] gonna feel it as well. I don’t know what it  is, but it’s there in all AI voices. It’s like   they follow a pattern that just… gives it away…  And once you notice it, you can’t unnotice it. So let me show you how I generate my voiceovers.  It takes much more time, but it’s the only way I  

[07:17] found to get rid of that uncanny feeling. And  it’s probably the secret sauce that makes you   wonder how this ginger dude sounds so realistic.  Alright, enough yapping, let’s get started. We have to choose a voice model first. V3 is  their latest model. It can do crazy stuff like  

[07:34] making the voice giggle, whisper, or be excited.  It sounds great, right? But when I tried it with   my own voice… “it sounded like this”. Who the hell  is that?! That’s like… my evil twin or something.

[07:48] I even tried it with other voices, and  every time, it slightly changes the voice   want. It’s still in the alpha stage, so hopefully  it will be more stable in the future versions.   But it’s not really usable right now. So I  use the V2 model instead for all my videos.

[08:04] Unlike V3, this model has a few more sliders.  Increase the speed, reduce the stability,   and… turn of that annoying music! I’m  trying to focus here, dude. Thank you.   Drag the similarity to something like  70%, and increase the style exaggeration.

[08:21] Now select just a few sentences from your  script, and paste it to Elevenlabs. What   did we needed for a realistic voice? Changing  tone and speed, pauses, emphasis, and writing   like a human. I assume you already nailed the  last part. So let’s focus on the other three.

[08:36] Using 3 dots makes the voice sound disappointed,  or confused. To make it more excited,   use exclamation marks. It always works.  If it doesn’t, use more exclamations! but you can try 3 dots or commas  to make them a bit more natural.

[08:52] Another example, this is how it  sounded like before making the changes. “Say my name. You’re Isaac. You’re goddamn right”

[09:05] Pretty normal. And this is how it  sounds like after making the changes. “Say my NAME! You’re ISAAC…  You’re GODDAMN RIGHT!!” you can generate it again a few more times.

[09:22] “you’re goddamn right. you’re  goddamn right. you’re goddamn right” But why are we generating a few sentences at a  time? Couldn’t we do the entire script at once?   Well, we could, but we shouldn’t. One reason  is, if it generates a bad take, we can easily  

[09:35] regenerate it without wasting too much credits.  But the second reason? This is the secret sauce. If you realized, each time you hit generate,  “the tone changes slightly. the tone changes   slightly. the tone changes slightly” even  though we input the same text. So, we’re  

[09:50] taking advantage of that. For each generation,  you get 2 extra regenerations for free. Use   those free generations, and download them as  well. You’re gonna see why in the editing part. With every batch we generate, we get a slightly  different tone and speed. That’s just like how  

[10:03] normal humans sound like. So when combined  together in the edit, it makes the final   result sound more natural and fixes that  uncanny feeling we talked about earlier. After knowing these, it just takes a few  hundred more generations, and you get this. Yes,  

[10:16] it takes ages, but it’s like cooking.  Small changes make a huge difference. Now put them all in a single folder, select them  all, and rename. That’s much more organized. Now we can move on to the next step.

[10:29] I’m gonna show this in Premiere Pro, but you can  do it in any editing software that has a timeline. best takes. Don’t worry about timing  or anything else at this point. Just  

[10:42] line the voiceover based on your script. If you  downloaded a few versions of the same sentence,   listen to all, and choose the  one that fits the context better. and combine them together. For this sentence, I  generated 3 versions. Let’s listen to each one.

[10:58] But the real magic, happens when you  chop some parts from each generation,   But the real magic, happens when you  chop some parts from each generation,   But the real magic, happens when you  chop some parts from each generation,  

[11:12] Now they sound good on their own,  but to make it more human-like,   I took the best sections from each  version, and got the final result. and combine them together.

[11:25] Once you got a single long straight line,  start cutting the pauses. Use the second   line to start the next sentence right after  the last one ends. But don’t forget to add   pauses and intentional spaces for  the important parts of your video.

[11:37] Now, even though Elevenlabs had a speed slider  in its editor, you may realize just now that   your voiceover is too slow, or too fast. Don’t  worry, it happens… I mean I’m not the only one,   right? Premiere Pro allows you to change the  speed, but it turns you into a chipmunk if you  

[11:52] do that. Oh, there’s an option to ‘keep  the pitch’. How’s that sound? Oh great,   it distorts the voice now. Is it really  that hard to make a speed changer, Adobe? Fine, I’ll do that myself…

[12:12] Great, the voiceover is almost ready. But we can  make it sound more professional. Here’s something   new I learned for this video. Take the parametric  equalizer effect, add a high-pass filter, increase   the h value, find echo-y frequencies, drag them  down, take this slider, find a point where…

[12:26] Or, just drag and drop the preset I made! I  also added a few fun voice effects and other   useful stuff as well for you to try out. So feel  free to download the link in the description. Well, I think we’re done here. But wait a minute…

[12:39] What does youtube say? Does it allow  using AI voices? Even more important,   Yes. What, you expected a long list of  explanations? Haha. I mean as long  

[12:54] as you’re making an original content, not funny  video compilations, yeah, you should be fine. just start using AI voices now? What  about, I don’t know… authenticity?

[13:08] Wow… that’s… better than I could ever do… See? Told you it would work. Sounds great, man!  Yeah, but… it’s just… not me. What’s wrong, man? You’re becoming  

[13:22] a youtuber. Wasn’t that what you wanted? I don’t know, it feels like cheating. It   feels more like I’m becoming a fraud, dude. A fraud? You wanna know what’s fraud?

[13:36] That… Silence… Having something worth saying, and  keeping it locked up because you’re too scared of   how it sounds. THAT’S fraud. Who says it should  sound like you? As long as it’s your own thoughts,  

[13:50] who cares how it sounds, huh? Come  on, we don’t have time for this. He was right. I was so obsessed with how I sound,   I was forgetting the most important  part. Actually saying something valuable…

[14:05] but that’s not all. I kept  the best one for the last… or if english isn’t your first language. I mean at least that’s progress,  

[14:22] but you’re thinking too small, dude. You’re  still thinking about replacement. Don’t you   see? Everything you learned, this isn’t just  about finding a voice. There’s a bigger picture.  Okay, name a big youtuber. I don’t know, does MrBeast count? 

[14:40] Too cliche, but okay. When did  his videos start to get viral?  I mean, he had over 10M subs in 2018. Yeah, but I mean, super viral. Look here. 2023. 

[14:52] Yeah that’s quite a jump. Now, when did he start dubbing his videos?  Mark Rober. Joe Hattab. Anyone who started dubbing  his videos to other languages, suddenly doubled,  

[15:06] tripled their views and subscribers. Does it really make that much difference?  Think about it. There are 2.5 billion active  youtube users in the world. Around half a billion   of them speaks english. That’s less than 20%. If  we just add up the next 5 most known languages  

[15:22] Do you see the potential? Sheet!

[15:35] Well, that was the twist. AI voiceover isn’t just  for narrating your scripts for you. It’s real   superpower is dubbing your videos. You can dub  your videos, and get a massive audience. To try   it myself, I added just 2 languages to my last  video, and the results were pretty clear… Oh,  

[15:51] it’s actually blurry, heheh. Don’t worry, I’ll  show it after showing how to dub your videos. As an example, let’s take my last video. I  already dubbed it into Spanish and Hindi,   so let’s do Portuguese now.

[16:04] Before we touch any AI, we need to separate  the ingredients. After finishing your video,   export as an MP3. Next, do the  opposite. Mute the voiceover,   and export an MP3 with only the music and  sound effects. You’ll see why in a second.

[16:21] Now go to ElevenLabs and click the ‘Dubbing’  tab. Create a new dub. Select your original   language, the target language, and upload that  voiceover we just exported. Hit generate, and   go make yourself a coffee. It takes a few minutes.

[16:34] Once it’s done, download the file and drag  it back into Premiere Pro. Here’s what you   this as a guide for timing. Middle,  is the new AI Portuguese voice. And  

[16:47] Line them up so the timing matches the visuals.  Once it sounds good, mute the original English   track, slap my preset on the Portuguese track to  make it sound crispy, and export just the audio.

[17:01] Now the only thing left, is to add it to your  video. In YouTube Studio, click ‘languages’,   and choose the video you just dubbed.  Make sure youtube didn’t automatically   dubbed to that language with their google  translate voice. If so, just delete it from  

[17:15] the menu here. Then click here, select the  dubbed language, and upload your audio file. Also, translate your video title and description  with ChatGPT to reach the 99.1% purity. And, you basically unlocked a new country!  Keep repeating it for other languages as well,  

[17:30] Now as promised, here’s what it  looked like in the analytics tab… So… yeah. It doesn’t look that impressive, does  it? I mean, where’s the millions of views? Well,  

[17:43] is how it looks like. Here’s Youtube’s  auto-dubbed Hindi track. And THIS,   And this is the difference in spanish dub.

[17:58] Wow… that’s… thousands of people…  who watched my video! People who   would have seen none of it if I had  listened to that voice in my head. Turns out, Steven was right…