How accurate is speech recognition in Polish?

Polish is recognized well by on-device engines. On 100 read FLEURS sentences, Whisper large-v3-turbo had a word error rate of 7.70 percent ¹ and Parakeet TDT v3 had 8.21 percent ². That is about one word in thirteen wrong, a good result for a language with seven cases and long inflected words.

Whisper is half a point more accurate; Parakeet is faster, at 19 times faster than real time ³ against 12 times faster than real time ⁴ for Whisper. Apple SpeechTranscriber has no row here: in our benchmark it ran only on English, German, French, Spanish, Italian and Portuguese, so for Polish the choice is between two engines.

Whisper and Parakeet are within half a point on Polish

With punctuation and capital letters counted as a reader sees them, Whisper's formatted error rate was 10.48 percent ⁵ and Parakeet's was 11.07 percent ⁶. The formatting adds close to three points for both engines, a larger step than in Dutch, so expect to touch up commas and capitals.

Published figures put Polish in the same range

NVIDIA reports 7.31 percent ⁷ for Parakeet TDT 0.6B v3 on the FLEURS Polish test, and 7.28 percent ⁸ on Multilingual LibriSpeech Polish. The Whisper paper gives 5.4 percent ⁹ for Whisper large-v2 on FLEURS Polish, lower than our 7.70 percent ¹ for the turbo model. Different model versions and a different set of recordings explain much of that gap, so read it as "Polish is a strong language for Whisper", not as a ranking of releases.

Polish ranks in the middle of the 25 languages, ahead of Czech

Polish is clearly easier than Czech (10.44 percent ¹⁰ with Parakeet) and Danish (14.57 percent ¹¹ with Whisper). It is close to Finnish (7.73 percent ¹² with Whisper) and Slovak (8.45 percent ¹³ with Parakeet). It trails Russian (4.59 percent ¹⁴) and the Romance languages, but not by a wide margin.

What to expect when you dictate in Polish

Expect mostly correct sentences. Typical errors are inflection endings, proper names and numbers. The error is a wrong word form rather than a wrong word, which still counts in the score but is usually easy to spot and fix as you read.

Choose Whisper if you want the lowest error on Polish and can wait longer for a transcript. Choose Parakeet if you dictate often and want text to appear quickly: the accuracy gap is small. Yap, which we make, uses Parakeet. Both engines ran fully on the device, so nothing was sent to a server.

How to read these numbers for Polish

A word error rate counts every substituted, deleted or inserted word against the written sentence. In Polish, a word with the wrong ending is a wrong word even though a human reader understands the sentence, so the score is stricter than the experience of reading the text. The formatted error rate adds punctuation and capitals on top, and that is the figure closest to what you see in a finished message. If you dictate in Polish and English in the same day, remember that English reached 3.92 percent ¹⁵ with Whisper in the same test.

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, Czech, Danish.