"Neural networks didn't understand Turkmen": How a developer from Turkmenistan is teaching AI to speak its native language

Bakhtiyar Mamedov, city — Kuala Lumpur, AI Developer at TakykAI

How It All Began

I started working with artificial intelligence back in the 10th grade — with recommendation systems. These are algorithms that suggest to users what they might like, such as on YouTube or marketplaces.

By graduation, I already had a solid knowledge base and several successful projects involving the fine-tuning of large language models — that is, “retraining” ready-made neural networks for specific tasks. However, such orders became extremely boring for me, and I began to think about what could truly interest me in the long term. That’s when I got involved in developing the Turkmen language in the field of AI.

At that time, there were simply no neural networks that understood the Turkmen language — neither text-based nor voice-based, meaning those capable of reading text aloud or recognizing speech. I started from scratch — with data collection. This project has been occupying me for two years now.

Initially, I invested not only a lot of time but also all my spare money into training the models. Any training of a neural network or its operation required renting expensive computing power — powerful graphics cards on which neural networks are calculated. Over time, as good results began to appear, people started noticing and offering collaboration. From several offers, I chose the most attractive one for the future project.

With the involvement of partners, growth accelerated several times over. We managed to assemble a team of interested developers, and finally, we officially launched.

The Biggest Challenge — Data

The main difficulty was not in creating the AI itself, but in collecting data. There was very little data, and it was scattered all over the internet. We had to write numerous programs to automatically collect information from websites and try different approaches to building models.

The speech models were the hardest to develop. There were no free or paid ready-made solutions that could recognize or reproduce Turkmen speech, and private studios charged large sums for recording hours-long audio materials.

Through crowdsourcing — when many different people work on a task bit by bit — we managed to train a speech recognition model. At first, it worked only on fragmented texts and audio and made mistakes in 10% of cases.

Since then, we have managed to reduce this error rate to 3%. Based on this model, we were able to collect data and train the first speech synthesis model — which converts text into spoken speech.

What We Have Achieved

Currently, we have some of the best language models in the world, a speech synthesis model, and a speech recognition model. This is also the first AI startup for the Turkmen language.

The main audience I targeted, contrary to all expectations, was small and medium-sized businesses. Only later came the users of our own application, similar to ChatGPT.

We attracted our first users through thematic Telegram chats and contextual advertising.

How AI Has Changed Work

Now, AI is the main tool in my development process. It speeds up idea verification and the creation of demo versions to just a few days or even hours. A year ago, I couldn’t trust a neural network with the entire development process and wrote critically important modules and everything related to security myself. But over the past year, AI has developed so much that I probably won’t be able to write better than it.

At the same time, I still manually review all the code generated by AI.

Plans

By the end of the year, we plan to complete integrations with B2B business clients, make the login process for businesses more convenient and faster, and also add paid subscriptions to the application.

In the future, I am looking at the Central Asian market. We are just beginning our research and evaluating how much different countries need their own national AI hosted on their own territory.

I have never had thoughts of closing the project. It has always been driven forward by the idea itself and the results already achieved.