When a small experiment outgrows its original purpose

I wanted to clone my own voice, but instead ended up with a project that somehow got 1,000+ GitHub stars.

All love stories

It all started in 2023. Initially, it wasn't even a "startup idea" or anything similar. I had only two use cases: I wanted to create videos for a developer blog without constantly recording myself, and I also thought it would be nice to create a voice assistant so my grandmother could hear me when I wasn't around.

Therefore, I tried to do it myself

I found a company doing voice cloning through a Habr article and reached out. I was ready to pay, just to use it. They told me they only do B2B.

The first line of the big journey

At the time, there weren’t many practical “plug and play” solutions. I looked at Tacotron2 and RTVC. RTVC didn’t really give me usable quality, so I started experimenting: changed parts of the pipeline, collected a dataset of my own voice, and trained models from scratch. It wasn’t clean or well-structured — mostly trial and error.

Back then, “10 seconds of audio to cloned voice” wasn’t really a thing yet, at least not in an easy form. So everything had to be built manually.

Eventually it worked well enough. I built the assistant and brought it to my grandmother. She listened for a bit and said she’d rather just hear me call her directly. Fair enough.

After that I kind of moved on, but stayed in the same general space. I got interested in animation and lip sync because the existing tools didn’t look good enough for the kind of videos I wanted to make.

The first users

That eventually turned into Wunjo. When I released it, things got messy pretty quickly. I built it on Ubuntu, and Windows users immediately ran into problems — driver issues, broken installs, all kinds of things I didn’t really anticipate. For a few months I was just going through GitHub issues one by one, trying to reproduce bugs and fix installs.

At some point I ended up rewriting the installation flow so the app could detect the environment, download dependencies, and pull models automatically on first launch. That part probably took longer than the actual model work.

No team, no budget — just users opening issues and me slowly fixing things.

The best way to create a product is to be your own customer. If you can’t be your own customer in full measure, there’s only observation left.

Now it has 1k+ GitHub GitHub, stars, all issues closed and around 15k reads on the first article . The “voice assistant for grandma” didn’t really become a product, but it kind of set everything in motion.