The most accurate Slovak speech to text we have measured
gaston-sk-v2 is built on KInIT's open Slovak Whisper model and further trained on about 2,600 hours of Slovak speech. It transcribes every Slovak recording in Gaston, from press conferences and parliament to podcasts and film.
Word error rate counts the words that differ from a careful human transcript. Lower is better.
Fewer errors than KInIT's open Slovak Whisper model
KInIT's whisper-large-v3-sk is an openly licensed Slovak model, and it is where we started. Further training on our own data cut its word errors by another 11.5% on average, and roughly halved those of the original OpenAI model.
It is better than KInIT's model on every one of the 7 short-clip test sets, each with 95% confidence.
Bar scale: 0 to 25%. Every model was measured by us, the same way, on test sets frozen before training.
Half the errors of OpenAI and ElevenLabs on parliament speech
We ran two of the best-known commercial transcription services, OpenAI gpt-transcribe and ElevenLabs Scribe v2, on the same six public test sets with the same scoring. On Slovak parliament speech our model makes about half their errors, and on average it makes fewer errors than either.
They are better on clean read sentences (FLEURS), and OpenAI also on Common Voice. We show where we lose too.
✓ gaston-sk-v2 is better ≈ tie within noise ▲ the other model is better. Bar scale: 0 to 25%. Tested in October 2026 on 7,605 public clips (about 16.4 hours) per service, with the language set to Slovak. A result counts as better only when the 95% confidence interval of the difference excludes zero. ElevenLabs writes filler sounds like "é" or "hm" that our references leave out; they were removed before scoring, so its numbers are its best case. Our own Gaston test sets were not sent to any outside service. Commercial models change over time.
Ahead on parliament, lectures and everyday speech
The biggest gains are where Slovak gets hard: crowd-recorded sentences and real recordings from Gaston users, the audio people actually upload.
On recent parliament speeches the model gets about one word in 70 wrong.
Bar scale: 0 to 25%. * KInIT's model was very likely trained on the parliament days we hold out from SloPal, so this row flatters it.
Measured on full files, not just short clips
Short clips are the standard test, but people upload whole interviews and broadcasts. So we also ran complete recordings through the real Gaston pipeline and compared them with transcripts a person proof-read word by word.
On these files the model makes about 24% fewer errors than OpenAI Whisper large-v3 and about 9% fewer than KInIT's model.
Bar scale: 0 to 25%. Real Gaston files that were not used for training, transcribed exactly as the production service transcribes them.
It keeps the surname
For journalists and researchers, a transcript without names is useless. A typical habit of KInIT's model is to keep the first name and drop the surname. Listen to this 18-second excerpt from an interview, which was not used in training.
The speaker says "Igor Matovič", a Slovak politician and former prime minister.
Už keď Igor Matovič dal taký status prednedávnom, dal tam výstrižok z relácie, kde bola Zora Jaurová a niečo ju kritizoval v tom zmysle, že ako môže sa, ako sa môže niekomu páčiť taký afektovaný prejav, takže to bolo také, ale hneď ju zároveň teda porovnával s vami.
Už keď Igor missing dal taký status prednedávnom, dal tam vystrižok z relácie, kde bola Zora Jaurová a niečo ju kritizoval v tom zmysle, že ako môže sa, ako sa môže niekomu páčiť taký afektovaný prejav. Takže to bolo také, ale zároveň ju teda porovnával s vami.
Už keď Igor Matovič dal taký status prednedávnom, dal tam vystrižok z relácie, kde bola Zora Jaurová a niečo ju kritizoval v tom zmysle, že ako sa môže niekomu páčiť taký afektovaný prejav. Hej, to ani niečo neviem. Takže to bolo také, ale popozorne ju zároveň teda porovnával s vami. Áno.
One example, picked to show the pattern. Amber marks spelling mistakes: "vystrižok" should be "výstrižok", and "popozorne" is not a word.
Counted on 11 real recordings (9.9 hours) that were not used in training, only where the rest of the sentence was the same in both transcripts.
It does not thank you for watching
Whisper models learned from subtitled video, so they like to add things nobody said: a closing "Ďakujem", a subtitle credit, a request to subscribe. We trained ours to stop.
Silence stays silent too. Long silent stretches come out as no text, where OpenAI Whisper large-v3 can drift into another language or repeat the same sentence over and over.
| Invented text | gaston-sk-v2 | KInIT | Whisper large-v3 |
|---|---|---|---|
| "Ďakujem" added at the end | 3 | 17 | 49 |
| "titulky", "odoberajte" or a web address | 0 | 1 | 30 |
Counted the same way for every model on 10,719 short test clips.
What went into it
We started from KInIT's open Slovak Whisper model, itself a fine-tune of OpenAI Whisper large-v3, and trained the whole model further on about 2,600 hours of Slovak: real audio from Gaston, parliament, film, lectures and read speech.
Clips of pure noise and music with empty transcripts teach it to write nothing instead of inventing text.
Our Slovak parliament dataset is now public: 471.5 hours in 140,420 clips from 190 members of parliament, November 2024 to September 2026, aligned to the checked official records of the National Council of the Slovak Republic. Free under CC BY 4.0, so anyone can train and test Slovak speech models on it.
What you get
Every Slovak recording in Gaston is transcribed with this model, in the app and through the Gaston API. There is nothing to switch on.
- •Colloquial forms are normalized to standard Slovak, for example "môžte" becomes "môžete" and "šak" becomes "však".
- •Czech speech is written as Slovak.
- •Rare foreign names can come out misspelled.
How we measured
- All test sets were frozen before training started, and none of them was used for training.
- Every model was decoded the same way (CTranslate2, fp16, beam search 5, language set to Slovak) and scored with the same text normalization.
- Numbers published elsewhere can differ because of other normalization and inference settings. Every number on this page was measured by us, the same way for all models.
- The gap to KInIT's model was checked with a paired bootstrap: on every short-clip set, the 95% confidence interval lies entirely in our favour.
| Test set | Notes |
|---|---|
Gaston, whole recordings Gaston |
76.5 minutes of real files, proof-read by a person, run through the full production pipeline |
Gaston sentences Gaston |
Real Gaston audio, corrected by a proof-reader |
TrePal NR SR 2024-2026 · CC BY 4.0 |
Slovak parliament, our own alignment of the official transcripts |
SloPal NR SR |
Slovak parliament, a few held-out sitting days |
VoxPopuli Meta |
European Parliament speeches |
TEDxSK TUKE |
Lectures |
FLEURS Google |
Read speech |
Common Voice 22 Mozilla |
Crowd-read sentences |
Credits
This model stands on open work. Thank you to everyone who built and shared it.
We trained the model to stop saying "Ďakujem". Here, we mean it.
- KInIT, whisper-large-v3-sk, Apache 2.0 (the model we started from)
- Based on OpenAI Whisper large-v3 (MIT)
- National Council of the Slovak Republic, official transcripts and recordings
- TrePal, CC BY 4.0 (our parliament dataset, built from those records)
- Meta VoxPopuli, CC0
- Mozilla Common Voice, CC0
- Google FLEURS, CC-BY-4.0
- MUSAN, CC BY 4.0 (noise and music for training)
- TEDxSK, TUKE
- SloPal (test only)
Try it on your own Slovak recordings
Upload a file or paste a YouTube link, and Gaston does the rest.