It started with a $5 game on itch.io.
The concept was simple and ridiculously fun: listen to a character, imitate the voice, and see how well you can match it.
I loved it.
The $5 wasn’t really the problem.
The problem was that I quickly ran out of things I wanted to play.
I wanted more packs～
For this kind of game, the mechanics are only half of the experience.
The packs are the game.
Once you’ve played the scenes you’re interested in, you immediately start wanting more:
More characters.
More movies.
More weird voices.
More scenes that make you think:
“Wait, I can totally do that voice.”
So I started looking for more material myself.
And that’s where things got ridiculous.
I was spending more time searching for scenes and preparing things I wanted to imitate than actually playing the game.
At some point I had a very developer-shaped thought:
Why am I spending all this time looking for packs when I could just build my own version?
So that’s what I did.
I AI-coded my own take on it
I started building Choicer Voicer, my own browser-based voice imitation game.
I used AI heavily while coding it.
Instead of spending weeks getting every small piece working from scratch, I could describe what I wanted, prototype it, test it, break it, fix it, and iterate extremely quickly.
The goal was simple:
Open a link → choose a pack → listen → record → get scored.
No installation.
No account.
No $5 purchase just to see whether you’ll enjoy it.
Just open the browser and start making questionable voice impressions.
But coding wasn’t actually the hard part
This was the unexpected part.
AI made building the software dramatically easier.
But it didn’t solve the reason I started the project in the first place:
I wanted good packs. Lots of them.
So after building the game, I ended up spending a huge amount of time creating, finding, preparing, testing, and reviewing pack content.
And I became surprisingly picky.
A clip being recognizable doesn’t automatically make it fun to imitate.
A good pack needs something more.
The voice needs character.
The delivery needs to be interesting.
The clip needs to be the right length.
The difficulty needs to feel achievable.
And most importantly, after getting your first score, you should immediately think:
“No. I can do better than that.”
I’ve now created and reviewed a large number of packs for the game.
Ironically, AI made the coding fast enough that content curation became one of the most time-consuming parts of building the product.
Then 200,000 people showed up
I originally built this because I wanted something I could play myself.
Then other people started finding it.
A few users became thousands.
Thousands became tens of thousands.
And now more than 200,000 people have played Choicer Voicer.
That was when I realized the problem wasn’t unique to me.
People didn’t just want the mechanic.
They wanted more things to try.
They wanted new packs.
They wanted to retry the same voice.
They wanted to compare scores.
They wanted to climb the leaderboard.
And, occasionally, they wanted proof that their terrible impression was at least slightly less terrible than their friend’s.
A score shouldn’t just be a number
This led to another problem.
Originally, finishing a recording could simply give you something like:
62%
Technically, that works.
But it’s boring.
If someone just performed an absurd movie impression into their microphone, the game responding with a sterile percentage feels like a missed opportunity.
So I started turning the result itself into entertainment.
Depending on your performance, the game can react like a tiny virtual movie set:
Director says:
“One more take!”
Crowd reaction:
“Wait… that was kinda good.”
And I’m experimenting with performance titles that change with your score.
The idea is that getting 95% should feel worth bragging about.
But getting 32% should potentially be even more worth sharing because the result is hilarious.
That’s something I didn’t understand when I started building the game:
Failure can be content too.
AI changed where I spend my time
Before this project, I would have expected most of my effort to go into engineering.
Instead, AI coding shifted the bottleneck.
I spend much more of my time thinking about:
What makes a good pack?
Which clips should I reject?
Where do players get bored?
What score makes someone retry?
What makes a bad result funny instead of frustrating?
What makes someone send their result to a friend?
Those aren’t primarily coding problems.
They’re product and taste problems.
And I suspect this is going to become increasingly common for independent developers using AI.
When implementation becomes dramatically cheaper, the question changes from:
“Can I build this?”
to:
“Can I make something people actually want to keep using?”
AI can generate a lot of code.
It can help me iterate incredibly quickly.
But it still can’t sit in my chair, play twenty packs, reject seventeen of them, and decide that one ridiculous three-second delivery is somehow the funniest thing in the entire game.
At least, that’s still my job.
You can try it
Choicer Voicer is free and runs directly in the browser:
Pick a pack, turn on your microphone, and see what score you get.
And if you run out of packs you like…
Well, that’s literally why I built this thing.
I’m still making and reviewing more of them.
If you try it, I’d love to know two things:
What score did you get?
And more importantly:
What pack should I make next?
⸻
Top comments (2)
One thing I forgot to mention: every pack has its own leaderboard.
Your best score is saved, and each pack has a Top 100 ranking.
The funny part is how quickly people get competitive 😅
Whenever I release a new pack, the Top 100 usually fills up within about an hour.
So if you see a brand-new pack… that’s probably your best chance to grab a spot before everyone else shows up 😂
choicervoicer.org/
You can also share your result directly with friends and challenge them to beat your score — which, unsurprisingly, makes people even more competitive 😂
Apparently all it takes is a voice impression, a score, and a leaderboard to turn your friends into rivals.