How I Made ToluWhispr

So basically, a few months ago, when I started getting into AI and everything, I found myself typing a lot.

And I mean a lot. Long, detailed prompts, instructions, all sorts of things. Sometimes I'd spend so much time typing something, forget to save it, and then have to type the whole thing all over again.

Now, one thing I'd noticed was that ChatGPT's transcription was amazing. I could just talk, and it would transcribe what I was saying almost perfectly.

But the problem was that I had to use it inside ChatGPT, then copy and paste the text wherever I actually needed it.

Then I discovered Wispr Flow.

And honestly, it felt like the solution to all my problems.

I could dictate globally on my laptop, in practically any application, and the transcription was really good. So I started using it.

The problem was that the free version was quite limited, and I wasn't willing to pay for it. It was expensive, and I didn't exactly have a lot of money to spare.

So the next thing I did was start looking for alternatives.

I tried Buzz, Handy, and a few other local alternatives. But my PC isn't particularly powerful, so running these models locally was just too slow.

I also tried the built-in Windows dictation tools, but they weren't nearly as accurate.

And that was the frustrating thing. I knew exactly what I wanted. I just couldn't find something that did it well enough without having to pay for another subscription.

Then I paid for my first ChatGPT Plus subscription and discovered that I could use Codex's dictation feature globally.

And it was fucking amazing.

In fact, that's what I'm using to dictate this right now.

But then my subscription expired, and I didn't renew it immediately, which meant I couldn't use the feature anymore.

So I was back to the same problem.

I tried a few more free alternatives, but none of them were quite as good.

Then, sometime later, ChatGPT randomly offered me another free Plus trial. I was actually surprised because I didn't even know why I was getting it.

Now, around the same time, I was scrolling through Twitter and came across this new speech-to-text model, Bark 3.0.

What caught my attention was how lightweight it was supposed to be, especially compared to the other local alternatives I'd tried.

And I thought, okay, this could actually be useful.

So I opened Codex and told it what I wanted to build.

I already knew the features I needed. I wanted something that could transcribe globally, run quickly on my laptop, save recordings, accept audio uploads, and just be easy to use.

The first version had quite a few problems.

It couldn't save recordings. I couldn't upload audio files. It was taking too long to launch.

And initially, it was just running through a batch file, but I wanted an actual application.

Then there was the global hotkey, which wasn't working properly. So we had to fix that too.

And there were a couple of other issues I can't even remember now.

So I kept going back to Codex, describing what wasn't working, getting it to make changes, testing them, and then going back again.

Now, when I say I was fixing these things, I wasn't actually writing the code. Codex was doing practically all the heavy lifting.

I don't know anything about app development, at least not enough to have built something like this myself.

But that's actually the amazing part.

I was able to go from having this problem to having something that worked in a surprisingly short amount of time.

And obviously, a lot of things had to be invented and developed for something like this to even be possible. I'm not ignoring that.

But it's still quite remarkable to me.

The app isn't particularly polished. There are things I could improve, and I'm sure an actual developer would find all sorts of problems with it.

But it works.

It's something I use, and it's genuinely made my life easier.

And I think that's what excites me most about all of this.

Not necessarily that I built some amazing application, because I didn't.

It's that I had a problem, I knew what I wanted, and I was able to make something that solved it without actually knowing how to develop software.

A few years ago, I probably wouldn't even have considered trying.

Now I can just open Codex and start figuring things out.