Slovak speech recognition is behind English. The reason is data, not models

Our goal in this project is to move Slovak speech recognition into the range where those applications become viable, and to release the resulting models openly so that anyone building for Slovak can use them.

How we approached it

We started by measuring. Before this work there was no way to answer which speech-to-text model performs best on Slovak, since public leaderboards evaluate on English or on multilingual sets where Slovak makes up a fraction of a percent of the data. We built the Slovak ASR Benchmark, now published on SLAIH: 39 hours of curated audio, stratified across domain, speech type and speaker gender so that no single source determines the result, with two thirds of it carrying added noise at several difficulty levels. That last choice makes our numbers deliberately harder than benchmarks run on clean recordings only, because real users speak from noisy places.

The benchmark quantified the gap we suspected. Across 22 evaluated models, the strongest commercial service reaches 10.8 % word error rate, while the best open-weight model available off the shelf reaches 17.5%. Close to seven points of difference, on the same data, under the same scoring.

That told us where to intervene. We ran a first round of fine-tuning experiments on Slovak data and the gains were large enough to reorder the leaderboard. Error rates dropped by roughly half across every architecture we tried, which is larger than the gap between the best open-weight model and the best commercial one.

So most of the work in this project is not modeling. It is collecting and curating Slovak audio with verified transcripts, across speakers, recording conditions and speaking styles. Verified transcripts are expensive, which is why we are also working on methods that let us train on audio without them. Untranscribed Slovak audio is far more plentiful, and being able to use it decides how far we can scale the training data. This is active work and it did not yet contribute to the models described below.

What we have so far

The first versions of our fine-tuned models are published on Hugging Face and listed in the SLAIH catalog. We selected the most widely used open speech recognition architectures that already handle Slovak reasonably well before any fine-tuning, namely the Whisper family, Nvidia Canary and Nvidia Parakeet. Starting from models that already have a usable Slovak representation means the fine-tuning refines something that works rather than teaching the language from scratch. These are also the models people are already running, so a fine-tuned checkpoint can be used in an existing setup without changes to the surrounding code. They were trained on all Slovak speech data available to us at the time, with audio augmentation applied during training so the models hold up on realistic recordings rather than only on clean studio audio.

The improvement on Slovak is substantial. We are also clear about what these first versions do not yet solve. Two limitations matter most.

Both are addressable and both are on our roadmap.

What comes next

We have collected considerably more Slovak audio since these models were trained, covering domains and recording conditions the current training set does not reach, and the weak labeling pipeline is intended to extend that further into audio we cannot afford to transcribe by hand. The next round of models will be trained on it and published in the same way.

If you have access to Slovak audio that could help improve open models, or a Slovak model you would like to see evaluated on the benchmark, we would be glad to hear from you: [email protected]