How accurate is speech recognition in Czech?

Czech is recognized with a word error rate of about 10 to 11 percent on a Mac. On 100 read FLEURS sentences, Parakeet TDT v3 had 10.44 percent ¹ and Whisper large-v3-turbo had 11.49 percent ². Parakeet is both the more accurate and the faster engine on Czech: 19 times faster than real time ³ against 14 times faster than real time ⁴ for Whisper.

About one word in ten is wrong. That puts Czech in the middle of the 25 languages, behind Polish and Slovak and ahead of Danish. Apple SpeechTranscriber has no row here: in our benchmark it ran only on English, German, French, Spanish, Italian and Portuguese, so for Czech the choice is between two engines.

Parakeet beats Whisper on Czech by about one point

Counting punctuation and capital letters as a reader sees them, Parakeet's formatted error rate was 12.75 percent ⁵ and Whisper's was 13.30 percent ⁶. The gap between the engines narrows from about one point on words to about half a point on formatted text, so for finished text the two are closer than the word figures suggest.

Published figures agree on roughly 11 to 13 percent

NVIDIA reports 11.01 percent ⁷ for Parakeet TDT 0.6B v3 on FLEURS Czech, close to our measurement. The Whisper paper lists 13.3 percent ⁸ for Whisper large-v2 on the same language, a little higher than our turbo model. Both published numbers match the order we found: Parakeet ahead, Whisper behind by one to two points.

Czech is harder than Polish and Slovak, its closest relatives

Czech and Slovak are closely related, yet Slovak scored better in our test with Parakeet: 8.45 percent ⁹ against 10.44 percent ¹ for Czech. Polish also scored lower, at 7.70 percent ¹⁰ with Whisper. Croatian is a near tie: it reached 11.34 percent ¹¹ with Whisper, so compare engine by engine on the dataset page. Czech is clearly easier than Hungarian (14.09 percent ¹² with Parakeet) and Danish (14.57 percent ¹³).

What to expect when you dictate in Czech

Expect about one correction per ten words: wrong case endings, names and numbers are the usual suspects. The text is usable as a first draft, and you will read through it once.

Parakeet is the better default for Czech: lower error and higher speed. Yap, which we make, uses Parakeet. Whisper is a reasonable second choice if you already run it. Czech is also among the languages in Claude Code's voice dictation list ¹⁴, which spans 20 languages from Czech to Ukrainian.

How to read these numbers for Czech

A word error rate counts every substituted, deleted or inserted word against the written sentence. Czech has seven cases, and a word with the wrong ending is a wrong word even when a reader understands it, so the score is stricter than the experience of reading the text. The formatted error rate adds punctuation and capitals, and it is the closest figure to what you see in a finished message. At this level the engines differ by a point or two, so your microphone and a quiet room will often matter as much as the engine you choose.

What the data does not show

The recordings are read sentences from Wikipedia, spoken in a quiet room, and every engine heard the same 100 of them on one Mac in one run. Real dictation has hesitations, names, numbers and background noise, and none of that is in this benchmark. Numbers and spelling variants are not normalized, so "25" against "twenty-five" counts as an error for every engine alike. Treat the gaps of a point or two between engines as small, and the gaps of five points or more as real.

Where to read more

• All languages in one place: dictation languages, accents and translation and the guide to speech to text in European languages.
• The full rows and every transcript: the on-device ASR benchmark, the published word error rates and the speed data.
• Sibling pages: Slovak, Polish, Danish.