How accurate is speech recognition in Croatian?

Croatian is usable but clearly harder than the big Western European languages: about 11 to 12 wrong words per 100 on read speech. In our October 2026 benchmark on a Mac with an M2 Pro, Whisper large-v3-turbo had a word error rate of 11.34 percent ¹ and Parakeet TDT 0.6B v3 had 12.44 percent ², on 100 Croatian recordings from the FLEURS test set. We have no Apple SpeechTranscriber result for Croatian.

Whisper is about a point better than Parakeet in Croatian

Whisper has the lower error rate, by roughly one point, but the gap is small on 100 recordings and the published figures below are closer still. Parakeet is faster: about 20 times faster than real time ³, against about 12 times for Whisper ⁴. Both are fast enough that you will not wait on either. The speed dataset covers every language.

Our measurement of the engines in Croatian

Lower is better in this table.

The transcripts behind each row are on the on-device benchmark page.

Published results sit in the same 11 to 13 percent band

NVIDIA reports 12.46 percent for Parakeet on FLEURS Croatian ⁵, within a few hundredths of our 12.44 percent. OpenAI's paper gives 13.4 percent for the older Whisper large-v2 ⁶, a little worse than the turbo model we ran. The agreement is reassuring: the Croatian error rate is a stable property of these models, not a quirk of our run.

All published rows are on the published accuracy page.

Croatian ranks sixteenth of 25 languages

Taking the best engine in each language, Croatian is sixteenth of the 25 languages in the benchmark. It is well behind Portuguese (3.28 percent with Whisper ⁷) and Ukrainian (5.65 percent ⁸), and well ahead of Slovenian (19.83 percent with Parakeet ⁹), its neighbor. Croatian falls in the middle of the field, with the Czech, Romanian and Greek results nearby.

What to expect when you dictate in Croatian

In a quiet room, expect roughly 1 wrong word in 9 with either engine, so plan to proofread every dictated passage. Errors are likely to land on names, numbers, rare words and word endings. Whisper is the first engine to try, and you should not expect Parakeet to fix what Whisper gets wrong, since their error rates are close. Speak in full phrases rather than single words, since a model needs context, and keep a custom vocabulary or correction list if your app offers one. We found no published fact that confirms Croatian as a built-in Mac or Windows dictation language, so check your system's language list.

How to choose an engine for Croatian

With error rates this close, the choice between Whisper and Parakeet is mostly about your app and your hardware. Whisper has the lower error in our test, and Parakeet is faster. If you proofread everything anyway, the speed gain is of little use, and the lower error is worth more. Dictate a page of your usual text through each engine you can run and count the corrections, since names and technical words will behave differently from the read sentences used here. The published accuracy page shows how other models compare on similar tests.

What these numbers do not show

The recordings are people reading prepared sentences in a quiet room, scored on one Mac with 100 recordings. Spontaneous speech, regional speech such as Kajkavian or Chakavian, noise and domain vocabulary are not in the test, and we did not measure Apple's model. Word error rate also counts a wrong word ending as a whole wrong word, which makes it strict for a language with many endings. For the measure itself, see the guide on how accurate speech recognition is.

Our guide to European languages in speech to text helps you choose an engine, and the languages overview covers other languages. For comparison, read the same question for Slovenian and Ukrainian.