Which languages does speech recognition handle best?

Among the 25 European languages we tested, Italian, Spanish and Portuguese are handled best, and Maltese, Slovenian and the Baltic languages worst. Taking the better of Parakeet TDT v3 and Whisper large-v3-turbo for each language, Italian reached a word error rate of 2.52 percent ¹ and Spanish 2.53 percent ². Maltese was the hardest, at 22.02 percent ³. That is roughly one wrong word in forty for the best case and more than one in five for the worst.

The ranking says what current engines do on read speech in each language, not what is possible. Every language here is European, so it does not cover Asian, Arabic or African languages. Each language had the same 100 recordings from the FLEURS test set, on one Mac. The full benchmark dataset lists all 56 data points ⁴.

Italian, Spanish and Portuguese lead at under four percent

Best result per language, either engine, word error rate in percent (lower is better):

Western European languages with large amounts of public speech data cluster at the top. English is in the top four, at 3.92 percent ⁶, rather than first. In every language above, a dictation draft needs only light correction. French (5.29 percent ⁹) and Dutch (6.32 percent ¹⁰) are in the next group down.

Maltese, Slovenian and the Baltic languages are hardest

Same measure, the six highest error rates among the better engines:

Maltese is the extreme case. Whisper's error rate on it was 72.35 percent ¹⁸, so Parakeet is the only usable choice there, and even it misses more than one word in five.

Greek shows that the engine can matter more than the language. Parakeet had 31.66 percent ¹⁹ in Greek, while Whisper had 11.96 percent ²⁰. A language that looks hard for one engine can be fine for another. See Whisper or Parakeet.

The two engines differ most in the smaller languages

Across all the languages each engine supports, the median error rate is 10.11 percent for Parakeet ²¹ and 11.34 percent for Whisper ²². The medians sit well above the best languages, so a typical language in this set is harder than Italian or Spanish.

Apple's SpeechTranscriber covers only six of the 25 languages, so we compare it separately in the Apple against Whisper question. Apple's on-device Dictation list covers 22 languages ²³, and built-in dictation features have a median of 15 languages across the products we tracked ²⁴. Language support is a separate question from accuracy: an app can list a language and still do it poorly.

What this means if you dictate in your language

• Top group (Italian, Spanish, Portuguese, English, Russian, German): expect a clean draft. The choice of engine matters little.
• Middle group (Dutch, French, Ukrainian, Polish, Czech, Slovak, Romanian): good results, with more corrections for names and rare words.
• Hard group (Baltic languages, Slovenian, Maltese, Danish, Estonian): useful for notes, but plan to proofread everything, and pick the engine for your language rather than the overall winner.

What the data does not show

• Read speech in a quiet room. FLEURS sentences are read aloud. Real dictation has names, jargon and noise.
• A small sample per language. Each score comes from 100 recordings by a limited set of speakers, so close ranks (a few tenths of a point) are ties.
• European languages only, three engines. Languages like Mandarin, Hindi or Arabic are not in this dataset. See is speech recognition worse in other languages.
• Word error rate counts words. Languages that form long compound words, or that write without spaces, are not comparable on this measure.

Related reading

Start with our overview of speech recognition and the guide to on-device speech recognition. For single languages see German and Spanish; for support lists see what languages Whisper supports. Speeds by engine are in the speed dataset.