Many people spend years waiting for the perfect job, team, or funding to start something ambitious. Saheed Azeez did not wait. While still an undergraduate at the University of Lagos, he set out to build a text-to-speech AI model that could speak in African-accented English. He named it YarnGPT, and trained it himself, at home, with personal savings, Google searches, and an unwavering drive to solve a problem he could see right in front of him: the lack of audio tools that speak like us.

The journey is still unfolding, but Saheed’s story already offers something powerful: a reminder that innovation does not always come from labs or startups. Sometimes, it comes from someone who is just curious enough to try.

We caught up with Saheed to talk about how it all started, the data struggles, and where he sees this work going next.

My name is Saheed Azeez Ayanniyi. I am a Mechanical Engineering undergraduate at the University of Lagos, currently in my final year. I got into programming by accident. One day I walked into a lab and saw people learning to code. I joined in, started with Python, and then later got into machine learning through competitions. That was how I entered the AI space.

Right now, most of the projects I work on involve audio data, but I would not say I am fixed in one area yet. I prefer Natural Language Processing (NLP) at the moment, but I see myself exploring more sides of AI in the future. Anything AI-related interests me.

Yes, clearly. It was in 2019 during a Python class. Someone mentioned machine learning and I got curious. At the time, I was studying mechanical engineering, so when I heard “machine learning,” I thought it was a part of mechanical engineering. That part "machine" just made me think that way.

From there, I found a Coursera course from IBM on data science, which I audited during COVID. That was my first structured learning in machine learning. I did not get a certificate, but it gave me the foundation I needed.

It started with seeing someone else do a text-to-speech project. The developer was Nigerian, but he built it for American English. He made the repository public, so I could look through it. That made me think: if he can do this for English, why not try it for our languages or accents?

Another reason was that I applied for a role at a company working on similar projects but did not get in. So I decided to just try building something similar on my own. In addition, I enjoy doing side projects for fun and for learning. Whenever I am at home and not busy, I like to pick a big challenge and try to solve it.

These kinds of projects also help recruiters find you. That has always been in the back of my mind: if I build something useful and share it, someone might notice.

The hardest part was data. That was the biggest challenge. If you want to build something for American English, you can find thousands of hours of audio data online. But for Nigerian-accented English? It is just not available at that scale. There are some initiatives like Naija Voices working on this, but it is still early.

To work around it, I used audio from Nigerian movies and matched them with subtitles. It was not perfect. There were some issues for example, the alignment between subtitles and audio was often off but I had to make it work. I did not even know much about cleaning audio data back then, so I mostly used it as-is. I was nervous about whether it would work at all, but the results were better than I expected.

The other challenge was technical. I had fine-tuned models before, but never trained one from scratch. I used PyTorch for the flexibility, but I had to learn a lot as I went. I relied heavily on blog posts, random articles, and lots of Googling. There was no single resource I followed. It was just solving one problem at a time.

And of course, cost. Training needs GPUs, and they are expensive. I was working at the time, so I used part of my salary from November and December to fund the project. I basically squeezed what I could from those two months to pay for it.

Yes, it was a solo project. I gathered the data, wrote the code and trained the model. It took me about two months from the idea to the release. I started in late November 2024 and launched the model on January 23rd this year.

The feedback has been great but also very real. A lot of people have said the model is not perfect yet, and they are right. It still makes errors and cannot be used in high-stakes settings just yet. But people also see the potential.

One person reached out and said they were using the model to generate audio for a cartoon they are building. That was a great use case to hear. I also see a future where it could be used in audiobooks, voice assistants, or customer service systems that speak in our local accents. But for that to happen, the model needs to be almost perfect.

Right now, I work with an organization where we are already building text-to-speech models, so my focus has shifted a bit. I do not want to say too much about YarnGPT 3 just yet, but the idea is not off the table.

The truth is, the feedback from YarnGPT 2 has been a lot to handle. It has kept me busy. People message me constantly about it. I have not had the time or mental space to start something new yet. But knowing myself, I might wake up tomorrow with a new idea and jump into it. That is how I operate.

So many. I have received several job offers. And not just any offers. Some of them I actually had to turn down because better ones came along. That kind of thing would not have happened before this project.

Beyond that, I have been in rooms I never would have imagined. Meetings, events, and conversations with people from companies I used to admire from a distance. That has been surreal.

Most importantly, it has made me more motivated. It reminded me that I should not limit myself. If I can dream it and work on it, I can build something that matters.

I love scrolling Twitter. I also watch movies and follow football. When my team loses, I enjoy watching other people get frustrated about it.

I also like resting. I take rest seriously. Once I finish a big task, I step back and chill. Even when working on projects like YarnGPT, there are stretches where you are just waiting for data to download or code to run. I use that time to relax.

I do not really go out unless I need to buy something. Most times, I just sit in my backyard and stare into space.

Elon Musk. He is not your typical CEO. He talks a lot in public, sometimes making controversial or unconventional statements, but he still delivers results. That is fascinating to me.

He has also faced serious challenges and kept going. Whether it is Twitter, SpaceX, or Tesla, he keeps solving problems on a global scale. I may not agree with him on everything, but I would love to understand how his mind works.

Start small. Be curious. Do not wait for everything to be perfect.

You do not need to know everything before you try. Just pick a problem you care about and start figuring it out. Use the internet. Read blogs. Break the problem down.

You will be surprised by how far you can go just by staying consistent. And do not be afraid to share what you build. That is how people find you.