Descript’s Text-to-Speech in 5 Easy Steps!
Ross Zeiger shows how to create custom AI voices in Descript by recording an authorization script, then generate speech by pasting text and converting it to an editable audio layer. You can add emotion tags like [excited] or [whisper] using ElevenLabs V3, and make word-level edits before finalizing.
Chapters
You can turn text into speech and make it sound like it's you speaking those words, and you can even control the emotion of that AI speech. Here's how it works. We're gonna start on the homepage, which is what you see when you first log into Descript, and we're gonna start by training the AI speaker on your voice. So come over to the left side where it says AI speaker, click on that. And you'll see here a list of all the AI speakers you've ever trained. If you haven't created one yet, there'll be a new speaker button right in the middle, or we can come up to the top right, click on new speaker, and this dialogue box pops up. Now what we're gonna have to do is give our speaker a label. So the way I recommend doing this is give it the name of the person who's training the voice. So I'll name it Ross. And then what I also do is I make videos in a lot of different locations. So I have a home office, I have a studio, I have different places I record. I also have different microphones. So I want to capture both of those environments. So I'll say Shure MV7+ Home Office. And then that just lets me know the environment that I train this AI speaker in for when I need to come and use it. So I'm going to hit confirm. And then what I'm going to do is read this script. And the script does two things. One, it authorizes Descript to use your voice. It's saying, yes, this is me. You have permission to train the AI on it. Data privacy purposes. The second purpose is it trains Descript on all the different sounds that it needs to replicate your voice. That means the quality of the input here is the quality of the output. So if you read it flat and monotone, it's going to come out even flatter and more monotone. It's going to sound very robotic. So I recommend over-enunciate, be more animated than you would normally be. Think 120%, 125%. Of your normal voice. And then to start it though, you're going to hit record, you're going to read it, and then when you're done, you'll hit stop. And before you do that, you want to make sure you have the right microphone selected. So I'm going to go ahead and hit record, and I'll meet you on the other side. I want Descript to create an artificial version of my voice. We tell stories, ask questions, and share jokes. And then it gives you this little box when you're done. Notice it took me 30 seconds to read it. You can re-record it if you don't like it, and you can preview it with this little button right here. And as soon as you hit submit, it's sending that off to Descript servers, and in a moment we'll have our voice ready to use. You can see at the bottom right, it says authorizing speaker. And there we go. It says speaker authorized. You can now use Ross Zeiger Shure MV7+ for speech generation. You can close that and meet me in a Descript project. Now that our speaker is trained and we're inside of our blank project, the easiest way to start is right here with this Create with AI Speaker button that appears right in the middle. Click on that and you can see you have a list of all the names you've trained on. I'm gonna find the one that I just trained. There it is. I click on it and I say choose speaker. And if you want, you can assign an avatar to it so you can upload a photo of yourself and it'll actually lip sync it so it looks like you're speaking it. That's beyond the scope of this video, but I have other tutorials about that on this channel and in my course. So I'm going to say skip this step. And now you can see it's labeled right there. And it's all, it's also going to be randomly color coded. So if you have multiple speakers, you'll have a green speaker, a blue speaker, a purple speaker, etc. And I can start writing something in here, or I can just copy and paste something. And it'll turn that into my AI voice. So I'm just going to go onto my computer. I'm going to copy something and I'm going to paste it in. And you can see it shows up as blue and you might get this little red pin here. What that is, is there's something unusual about what you copied in. It might be like an emoji or something that's non-text and it just shows up as a little red emoji. You can delete that or ignore it. So this light blue text means it hasn't been It's in there, but you can still edit the text. It hasn't been converted to audio yet. So for example, if I want to change gonna to going to, I can just click on it and type the correct spelling. And then I can come up to done writing. And actually, let me just make this bigger for you and click done writing. And then it's going to say generating AI speech. Okay, and there we go. So let's listen to this. Notice it turns a darker shade of blue. Once you've created it, that's telling me that this is an AI voice. I can hit play. Even flatter and more monotone, it's gonna sound very robotic. So I recommend over-enunciate, be more animated than you would normally be. So there we go. It sounds sort of like me. And now I wanna point out something about this AI text. Notice if I come down to the timeline, it's blue down here as well, and I'm unable to grab letters and move things around. I'm unable to trim this clip if I can't get the bracket on the either ends. What I do have instead is this button that says convert to audio. So until I click that, I can still make changes to my blue text. I can still change the voice here. So I do that by coming back to the write button in the top right corner. And then now we're back in editing mode. So I could click somewhere and I could just add in a word like the word and, and once I do that, it's going to get rid of the previous voice generation and I'm going to have to hit done writing again. So I click on that and now it's gonna regenerate my AI voice and it's done. But as before, as I started to point out down here, it's still in what I call draft mode until you hit this button that says convert to audio. So if you're happy with this generation, then we can click on convert to audio. And now notice everything turns black, which is the same color it is when you record your voice like I am doing right now. And now we have all the capabilities of editing so I can drag words around down on the timeline. I can trim the clip. I can highlight words here and delete them from my video and it's treated like a real audio layer. What that also means is now if I go into write mode, I'm not gonna be able to correct it like I could before. I'm not gonna be able to regenerate it. So be aware of that. And if I'd wanted to change the speaker while it was still blue, I could have clicked on my name and changed to another voice and then it would reprocess that text. In that other voice. On that note, let's talk about the stock speakers and changing emotions. So I'm going to start a new composition. I'm going to go to Create with AI Speaker. I'm going to scroll down past my own voices and I'm going to choose one of these stock speakers. So they have names like Alex, Bernard, Bradley, and it tells you what they sound like. So conversational, adult, inspiring, masculine. And you can do a little preview by clicking the play button next to their name. But I'm just going to go with Alex. It's at the top of the list here. So I'll click on it and then I'll say choose speaker. I'm going to skip over this avatar and I'm going to paste in the same text that I had before. Now this time I'm going to give it emotion. And the way I do that is with square brackets. So I type in a square bracket and I type in something like excited. You can also type in scared, whispering, things like that. And what the AI is going to do is it's going to not read the emotion or whatever is in the square bracket, but it's going to take on that emotion for the words that follow it until you give it another emotion. So after where it says, think 120%, or before it rather, I'm going to type in another square bracket and I'm going to say whispers. And then once I've got my emotions in there, I hit done writing. And as before, it's going to process that into the speech. And let's go back to the beginning and hear how this sounds.
Even flatter and more monotone. It's going to sound very robotic. So I recommend over-enunciate, be more animated than you would normally be. Think 120%, 125% of you.
Okay. Well, it did the excited pretty well. It didn't do the whisper so well. So your results may vary. And one other thing, one last thing to tell you about with this is if it's not recognizing the emotions, scroll down to where it says AI speech generation model and make sure you're on ElevenLabs V3. It only works on ElevenLabs V3. And if you hover over the I here, it says some stock speakers use a different provider. So not all the stock speakers will recognize the emotions. So be aware of that. Let me know in the comments. What kind of projects are you using the AI Speaker for? How do you plan to use what you learned here today?
