No one recorded a single word of that opening. I wrote a scene to try to break Drama 3, Fish Audio's new voice model. It has three characters, a fight, a whispered confession, and a really bad joke at the worst possible moment. Then I had Claude Code build a tool that turns any script into a full cast of voices, and threw the scene at it.
For a limited time, developers get 80% off Drama 3 on the API, so it's $10 for a million characters. All creator plans are 50% off right now, and if you sign up through my link you get 10,000 web app credits.
⏱️ CHAPTERS 00:00 No one recorded a word of that 00:33 From one AI voice to a full cast 01:48 The scene: $4,200 missing from the food truck 02:14 Six traps AI voices fall apart on 03:42 How Drama 3 tags work 04:45 Building the tool with Claude Code 07:41 Playing the whole scene 08:47 Checking every trap 10:02 No recording booth needed 10:30 How you could make money with this 10:48 A radio ad for a coffee shop 12:06 The same ad in Spanish 12:40 Side hustle ideas 13:40 Try it yourself
Comment what you'd build with this. And if you're able to break it, I'd like to hear about it.
$4,200, Marcus.
It was me.
So no one actually recorded a single word of that.
So I wrote a scene specifically to try
and break Fish Audio's new voice model, Drama 3.
I put three characters, a fight, a whispered confession,
and a really bad joke at the worst possible moment.
I did whatever I could to try and break this.
So today I'm going to build a tool
that turns any script into a full cast of voices,
throw this scene at it, and we'll see how it does.
Fish Audio sponsored this video.
Really cool tool.
excited to dive into it
so you guys know i have 25 plus agents that run my life and i wanted to
turn it more into Jarvis where i don't type so much
but i actually can just speak to it i think
it's cool in the Iron Man movies
and i want that for my life i think i might see myself a little
bit as Tony Stark without the bad morals
and billions of dollars so a couple of months ago
i used Fish Audio to give my ai assistant a voice
and that was one voice reading one thing in one
tone, which that's kind of the easy version.
But the version that I think is actually more useful
for a side hustle or for making money
is when there's more than one person talking. So you can
think about like an audio book with a bunch of different characters.
You can do a story channel
on YouTube, radio ads with two people speaking to each other.
And in the good old days, which again
is what I call pre-AI days, if you wanted to do any of those things,
you had to find voice actors,
get everybody's audio to match up
and then paying again anytime that you needed to make a change or
you changed a line here or there
but if one person on a computer with ai can do the whole cast that
opens up a lot of small jobs that you wouldn't have
been able to do before so this only works
if it sounds like real people if it sounds like
a robot reading a script nobody's paying for that
so that's part of what i'm going to be testing today
so here's the imaginary scene i put together
there's three friends that run a food truck together Dana Marcus
and Lily and four thousand
two hundred dollars just went missing from the truck account disaster
so i packed this with the
stuff AI voices usually fall apart on I put
in like all the traditionally hard things for AI to
replicate so these are the traps that I've seen
in AI voice generation in the past but trap number
one Marcus gets cut off in the middle of a sentence
so his line kind of just stops and a
lot of text-to-speech reads every sentence like it's finishing
a book report and it's all nice
and polite but I want to hear
if it actually sounds like he got interrupted so that's the
first thing I'm checking.
Trap number two Marcus does an impression of Dana so the same character
but for one sentence he's going to do like a whiny version of her voice
and then drop back
into his normal voice. So that's hard even for a real actor to do.
So the third thing I want to
test is Dana starts out really angry. She's furious.
And then a few lines later, that anger
kind of drains out and she's tired.
So the same person, but totally different energy. So I want
to see how AI handles that situation.
Number four, this is a big one. I'm really curious to see how
AI handles this, but Lily's been quiet this whole time.
And then she whispers, it was me. Okay. So
how to whispering is really hard for AI to do.
It sounds really funny. If you've ever asked ChatGPT
to whisper something to you, it sounds very, very funny.
Fifth thing I wanted to test was sarcasm
right before the confession. Sarcasm is all timing.
If it reads that line straight, then it
doesn't make any sense. Sarcasm is all about your tone.
And number six, Dana laughs in the middle of
a line and then gets gentle in the same breath.
So let's see how it does with that. Then also I
put a dollar amount in there
because numbers trip AI up very badly traditionally so I want to see
how it does with that so looking
at Fish Audio's new model Drama 3 what's different about this is
you don't pick emotions from a drop down like you've done
in a lot of other voice generators
you just write what you tell an actor in brackets right in the text
so it's starting to what we send
to AI is pretty much what we would send to a human actor
and again seeing if AI can actually
replicate what a human actor would
do so before Dana's line I just put in brackets furious talking
fast but like for Lily I put barely above a whisper scared slow okay
so and you can do multiple
in the same sentence because like we don't usually have
the same tone through a whole sentence
but like for Marcus I put mimicking her high
and whiny and then later flat his normal voice
so then you can also use tags like in a sentence
so for Lily I put a tag whisper these are the
only words I want her to whisper
so with Drama Three each character gets its own voice so I can
do several voices in one request I don't have
to like generate each person individually I can just
do this whole conversation
so we have the script written out I have Fish Audio ready to go now we
want to build the thing that turns this text into audio
and I'm not writing a Python script by hand
or anything like that I'm gonna have Claude Code build it
as it does everything else in my life
at this point. And I'll kind of show you
the parts that really matter. So I already signed
up for Fish Audio. I grabbed an API key.
API key basically just lets you call it from your own
scripts. So that's all set up.
So the prompt that I'm actually giving to Claude Code, I'm just
telling it to read the Fish Audio text-to-speech docs first.
And this is a tip that I've learned
about AI is that if you're using outside tools,
you can have AI go read all the docs. So then it
knows exactly how to communicate with it.
So this is actually, again, better than a human
because traditionally as a programmer,
if I'm going to go access an API, I might not check the
docs at all. I'll just kind of hope that I get it right.
Or I might try to find an example someone
else did. But I never have read the entire docs of any API.
But this is an example of where we can
have AI go do what you really should do.
If you want to use the tool properly, go read all the
docs so I can have Claude Code first go read all
the docs for the Fish Audio API then it's going
to use it exactly as Fish Audio wants me
to then I'm having to write a Python script so again the
beauty of AI I don't really know Python
that well I don't use it on a day-to-day basis but I can
have Claude write a Python script I'm going
to call it just scene.py takes a script file looks
through it gives each character their own speaker number
in the order they appear then it's going
to get the some voice IDs from a file I'll show you here
in a second send the whole script as one
request to Fish Audio and then get back an audio file
and then also the captions that'll be handy
as well so I'm going to get it running here here's
the exact prompt that I'm using and actually
Claude even helped me generate the prompt
so we have total AI-ception we have AI writing the
prompts for AI to contact AI. It's just AI all over.
So you can see it's first going to the docs
for Fish Audio and looking exactly how it should contact it
and the exact commands that it should
give. So we'll give this a couple minutes here
and we'll see what comes out. All right, took three
minutes and 23 seconds. Not bad at all.
So here is the Python script that it wrote. So it's using
the drama three model but it parses
the script puts it all in the right format to send to the
API and then sends it to the Fish API
so my goal is that i can just write any script like this
send it to this script that i just wrote
and it should just create a conversation so let's go
ahead and run this time python three scene dot py scene dot
so it has the three speaker IDs so
Let's go ahead and open this up.
We'll open up QuickTime Player.
Scene one, take one, and let's listen.
$4,200, Marcus.
$4,200 is gone from the truck account,
and you want to talk about the menu?
I'm not talking about the menu.
I'm talking about why you're yelling at me in a parking lot,
because I never even...
You're the only one with the card.
You're the only one with the card.
Yeah, because you lost yours.
I'm not blaming you.
I'm asking.
There's a difference.
It was me.
What?
I used it for my mom's rent.
I was going to put it back before anybody looked.
I'm so sorry.
Well, good news, Dana.
I'm not a thief.
I'm just the guy who drove four hours for a fryer.
You're unbelievable.
Hey, Lily, look at me.
We'll figure it out together.
Isn't that a touching scene that I put together?
So let's go through some of the traps one by one.
I'm not talking about the menu.
I'm talking about why you're yelling at me in a parking lot
because I never even.
You're the only one with the card.
So that's the cut off.
I think it's pretty good.
There's like a tiny delay more than what I would expect.
And, you know, it sounds more like
a voice actor than it does like an actual conversation.
Like this sounds like a movie, right?
so second thing i was testing was the impression you're
the only one with the card and i think
that sounds pretty good well if i took out
this barely above a whisper scared slow um that's
pretty good direction of like how i want it to sound
but if i just change it to sad then i can
generate it again it was me but that's what you get
when you just say sad uh it just it's not as
good as if you give it more direction of okay how exactly
do you want it to sound i can change it to
barely a breath like she's afraid the others will hear
if you're a good writer then you can do some
cool stuff with this voice shaking she's
been holding this in for days it's eating her up it
was me see the difference there though uh
so this is fun you can uh really mess around with this
so three voices one script no one in a recording booth did all
this i mean how long did this take
When you watched me, you know, five, 10 minutes,
three minutes to write the script, three minutes
to write the Python script in Claude code,
which I again was just a command. Honestly, the hardest
part now is writing the script that you want it to actually say,
and then figuring out what to do
with it. So, and it did, it wrote the captions
to go along with it. So if you're doing, you know,
something on YouTube where you need to upload the captions,
it puts it with the timestamps exactly
matched up with the audio
so how do you monetize this kind of power now that i am have a million
an unlimited number of voice actors at my disposal
so let's brainstorm a little bit how you could
actually use this to make something
that you could sell how can you start a side hustle from
something like this first off i put together an ad for a coffee shop
so two people a barista
and a customer who's half asleep
so same tool i'm going to just use this now to turn this
into an audio file morning what can i get you whatever has
the most coffee in it then you want
the day break two shots oat milk
and a cinnamon thing we don't have a name for yet okay wow that's
actually great Corner Cup open at six on Main Street
so we got a radio ad that you can put
together here and then and the beauty of this also
is maybe you send this to a coffee shop and the
owner gets back to you and says yeah i really like it
but can the barista sound more excited
about the new drink so we can just go into the script
and instead of just saying laughing
let's say and genuinely excited about
this drink like she invented it and it's the best
drink ever ever made in history
so then we can open it take two then you want the day break
two shots oat milk and a cinnamon thing we don't have a name for yet
so i changed one direction
ran it again that kind of revision traditionally would take
a booking the voice actor again
getting them to say the line again so we can
also do crazy stuff like one more ask let's say the
owner wants it actually in spanish
for their spanish-speaking customers i actually just had
Claude Code translate the exact same script
into spanish buenos dias que te sirvo lo que tenga
más café entonces quieres el daybreak doble espresso leche de avena
y una cosa de canela
que todavía no tiene nombre okay wow está buenísimo Corner Cup abrimos
a las seis en Main Street
there you go i'd love for you to use
your creativity post in the comments what do you
want to do with this but if i were going to turn
this into a side hustle um ai keeps telling me to
stop starting new side hustles.
But if you want me to run with this one, I will, because it sounds
interesting. But one, you could do short audio dramas
for a YouTube story channel. You write the
scene, cast the voices, the captions come with it automatically.
Do some radio ads, do some social
ads for local businesses, you know,
plumbers or dentists or whatever, the coffee shop down the
street. I think people will be interested in this.
And again, they could do it themselves,
but people are busy and they're not going
to so if you're the AI person that can do this it's
probably even fine to tell them you're going to use AI
but they don't know how to do this you
could tell them to go subscribe to the James Pelton channel
and then they might figure it out
for themselves but you could send this to people you
do picture books bedtime stories with a
different voice for each character very easily
because it's an API you can plug it in to anything
a website an app a spreadsheet full
of scripts i mean whatever you're whatever you want to do
So if you want to try this for yourself,
start with the Fish Audio API, set the model to Drama
Three, right? The hardest scene you can think of.
And let me comment some of the results I want to
see. Developers get 80% off Drama Three via API.
So it's just $10 for a million characters. You get
a lot for a million characters is way cheaper than image generation
and way cheaper than video
generation. And all creator plans are 50% off right now.
So if you register through my link,
you'll get 10,000 web app credits that you can use.
So link is in the description.
And if you are able to break it, I'd like to hear about it.
And honestly, they would too,
because they're constantly just improving their models
and getting better and better, so that feedback is great.
So shout out to Fish Audio for building something cool.
This is one that has been a lot of fun
for me to mess around with.
Please do like and subscribe,
comment what you want to do with this.
Appreciate you guys watching.
God bless you all, and have a great rest of your day.