How accurate is speech recognition in Portuguese?

Portuguese is one of the easiest languages for on-device speech recognition: the two open engines make about 3 to 4 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 3.28 percent ¹ and Parakeet TDT 0.6B v3 had 3.62 percent ², on 100 Portuguese recordings from the FLEURS test set. Apple SpeechTranscriber scored 100.00 percent ³, which means every word was counted wrong.

Whisper and Parakeet are close, and Whisper is slightly ahead

The gap between the two open engines is about a third of a point, which is small on a sample of 100 recordings. Parakeet is faster: about 22 times faster than real time ⁴, against about 16 times for Whisper ⁵. Both are far faster than speech, so for dictation the speed difference rarely matters. The speed dataset shows every engine and language.

Apple's 100 percent result is a failed run, not a verdict

A word error rate of 100 percent is what you get when the output has no correct words, for example an empty transcript. Apple's model finished about 51177 times faster than real time ⁶, which fits a run that returned almost nothing. Our data does not say why, and it does not show that Apple's Portuguese is bad: we measured Apple in only six languages, and in the other five it came within a few points of the open engines. Apple's Dictation list for the Mac includes both Portuguese (Brazil) and Portuguese (Portugal) ⁷, but the on-device list has Portuguese (Brazil) and not Portuguese (Portugal) ⁸. Treat the row as "no result" until you try it on your own Mac.

Our measurement of all engines in Portuguese

Lower is better in this table.

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

Published results put cloud models in the same range

On FLEURS Portuguese, the Voxtral paper reports 3.54 percent for GPT-4o mini Transcribe ⁹, 3.57 percent for Voxtral Mini Transcribe ¹⁰, 3.90 percent for Whisper large-v3 ¹¹ and 4.50 percent for ElevenLabs Scribe ¹². NVIDIA reports 4.76 percent for Parakeet ¹³, a little above our 3.62 percent.

The test set matters more than the engine. On Multilingual LibriSpeech Portuguese, Parakeet had 7.50 percent ¹⁴ and ElevenLabs Scribe had 15.40 percent ¹⁵. Different recordings, different text rules and different scoring give different numbers, so compare engines only inside one test.

Showing 8 of 27 rows. See every row in the dataset.

All published rows are on the published accuracy page.

Portuguese ranks third of 25 languages

Taking the best engine in each language, Portuguese is third of the 25 languages in the benchmark, behind Italian and Spanish and ahead of English (3.92 percent with Whisper ²⁰). Italian (2.52 percent ²¹) and Spanish (2.53 percent ²²) are a bit better. The Apple row is left out of this ranking.

What to expect when you dictate in Portuguese

In a quiet room with a decent microphone, expect roughly 3 to 4 wrong words per 100 from Whisper or Parakeet, which is a light editing pass. Pick an engine that you can confirm is running: if a built-in tool returns nothing or garbage, switch to Whisper or Parakeet rather than assuming the language is hard. Choose the variant you speak, Brazilian or European, since the Mac's on-device list names only Brazil. Our page on Brazilian and European Portuguese dictation goes into that choice.

What these numbers do not show

The recordings are people reading prepared sentences in a quiet room, scored on one Mac with 100 recordings. We do not know the accent mix of the speakers. Spontaneous speech, noise, names and technical vocabulary are not tested, and the Apple failure is unexplained. 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 Italian and Spanish.