Estonia builds its own open weights, and they work

"Tere, kas sa kuuled mind?"

That's about as simple as Estonian gets. Can you hear me. I ran that recording through OpenAI's Whisper today and got back "Tereha kassugulab mind."

That's not a near miss. Those aren't words.

I've built a second brain called Morpheus. It keeps my open loops, my notes and my projects, and this week I taught it Estonian. I talk to it from my phone and it answers out loud.

That's two separate jobs: hearing me, and speaking back. Estonia has an open model for each of them.

For hearing, there's TalTechNLP's Estonian finetune of Whisper. I ran real recordings from this week through both it and the general Whisper. Both running locally on a Steam Machine in my house, same audio, same CPU, and both told the language was Estonian, so neither one was handicapped.

Here's what I said, and what the general model heard.

The Estonian finetune got every one of them right.

For speaking, there's TartuNLP's neurokõne, from the University of Tartu. That's the voice Morpheus answers in, and it's free for anyone to use at neurokone.ee. They also released 92.4 hours of recorded Estonian speech openly, which is the part that lets someone else build the next one.

To be fair about what this is: a handful of clips, one speaker, one microphone. It's an illustration, not a benchmark. For real numbers there's Olev and Alumäe, "Open source platform for Estonian speech transcription" in Language Resources and Evaluation (2025).

But that's the whole point. A model trained on everything doesn't fail politely on a small language, it fails confidently. The only thing that fixes it is a local lab that builds for the language and then gives the weights away.

Both of these are MIT licensed and free to download. So, genuinely, thank you:

• TalTechNLP, the Laboratory of Language Technology at TalTech, Tallinn University of Technology
• TartuNLP at the University of Tartu

Aitäh.