Dictation Punctuation and Capitalization Accuracy

Dictated text needs a noticeable amount of editing in a typical language, and the two open engines are nearly tied on that. Across the 25 languages we tested, Parakeet TDT v3 has a median of 12.75 percent ¹ and Whisper large-v3-turbo a median of 12.94 percent ². Apple's SpeechTranscriber has a median of 10.62 percent ³, but it covers only six languages, and one of them failed completely.

The figures come from our October 2026 run on FLEURS, with 100 recordings per language. A figure is the number of errors per 100 tokens when the text is compared with the written sentence as a reader sees it, so a missing comma or a lowercase name counts. Lower is better. The sections below show where that editing load falls, how the engines differ, and what the data cannot tell you. For the same engines judged on words alone, see our on-device speech recognition benchmark, and for background on how these systems work, our overview of speech recognition.

The language you speak decides the editing load more than the engine

The spread between languages is much wider than the spread between engines. Whisper's figures run from 5.68 percent in Spanish ⁴ to 71.96 percent in Maltese ⁵. Parakeet's run from 5.44 percent in Spanish ⁶ to 34.83 percent in Greek ⁷. The same engine can need several times more editing in one language than in another, while the two engines rarely differ by more than a few points in the same language.

Three tiers show up in the rows.

• Light editing (the best languages). Spanish is lowest for both engines. Italian follows at 6.40 percent for Parakeet ⁸, and Russian at 6.13 percent for Whisper ⁹. German sits at 7.06 percent for Parakeet ¹⁰ and 7.10 percent for Whisper ¹¹. Portuguese is close behind, with 7.45 percent for Whisper ¹².
• Moderate editing (the middle). English is here, not at the top: 8.18 percent for Parakeet ¹³ and 7.71 percent for Whisper ¹⁴. So are French, Dutch and the West Slavic languages. Czech reaches 12.75 percent for Parakeet ¹⁵, and Swedish 12.94 percent for Whisper ¹⁶.
• Heavy editing (the hard languages). The Baltic languages, Slovenian, Hungarian, Greek and Maltese sit here. Lithuanian costs Whisper 26.71 percent ¹⁷, Slovenian 25.26 percent ¹⁸, and Estonian 20.52 percent ¹⁹.

The pattern follows how much each language has in the engines' training data and how regular its writing is. Large languages with plentiful speech and text come out best. Smaller languages with rich word endings come out worst, because a wrong ending is a word error before punctuation is even considered.

Neither open engine wins overall, and each has its own weak spots

The medians are 12.75 and 12.94 percent, a gap too small to act on. The language-by-language picture is more useful, because each engine wins in some places and loses badly in others.

Parakeet is clearly better in the Baltic and some Central European languages. In Lithuanian it reaches 20.07 percent ²⁰ against Whisper's 26.71 percent. In Estonian it gets 15.29 percent ²¹ against Whisper's 20.52 percent. In Slovak it gets 10.29 percent ²² against 12.58 percent ²³. In Hungarian it gets 16.56 percent ²⁴ against 19.28 percent ²⁵.

Whisper is better in the Nordic languages and in Greek. Finnish costs it 9.88 percent ²⁶ against Parakeet's 12.10 percent ²⁷. Danish costs it 16.90 percent ²⁸ against 18.54 percent ²⁹.

The two largest gaps are the most important rows in the set:

• Greek. Whisper needs 16.10 percent ³⁰ editing, Parakeet 34.83 percent ⁷. If you dictate Greek, the choice of engine matters a great deal.
• Maltese. Parakeet reaches 25.96 percent ³¹. Whisper reaches 71.96 percent ⁵, which means most of its output would be rewritten. Whisper's median is the one most exposed to this outlier, and its range is wide as a result.

Medians hide this. Two engines with near-identical medians can still be far apart for the one language you need.

Punctuation and capitals add a visible layer on top of word errors

The table shows, for each row, the formatted figure and the words-only figure (the column labeled Words only (WER)). The formatted figure is almost always higher, and how much higher depends on the language and the engine.

In the languages where recognition is already accurate, the formatting layer is a large share of what is left to fix. Spanish and Italian show it clearly: Whisper's Italian figure is 7.65 percent ³², yet its words-only figure is a fraction of that. In the hardest languages the opposite holds. Estonian, Latvian, Lithuanian and Slovenian have formatted and words-only figures that sit close together, because wrong words dominate and punctuation barely adds to them. For Whisper in Maltese the formatted figure is even marginally below the words-only one, which is possible because a token that is wrong either way is counted once.

The practical reading is that formatting is the main remaining edit in the strongest languages, while recognition itself is the main edit in the weakest ones. Better punctuation will help a Spanish or German dictator. It will do little for a Lithuanian one.

Showing 8 of 25 rows. See every row in the dataset.

Apple's engine is steady where it works, and failed outright in Portuguese

Apple SpeechTranscriber supports six of the 25 languages, and the five that produced usable text land in a narrow band. French is 10.34 percent ³⁷, Spanish 10.43 percent ³⁸, German 10.55 percent ³⁹, Italian 10.68 percent ⁴⁰ and English 11.10 percent ⁴¹.

Portuguese scored 100.00 percent ⁴², the same as its words-only score, which is the signature of an engine that returned no usable text rather than text with mistakes. We did not establish why. It is why Apple's median of 10.62 percent ³ sits near its best rows even though its range ends at the maximum. The median resists one failed language, but it should not be read as a typical Portuguese result.

Apple's figures are also flatter than the others. Its formatted figures are nearly the same across English, Spanish, German, French and Italian, while the open engines vary more. In the languages it shares with them, Apple is behind in Spanish and Italian: Parakeet gets 5.44 percent in Spanish ⁶ and 6.40 percent in Italian ⁸, against Apple's 10.43 percent ³⁸ and 10.68 percent ⁴⁰. In French the gap narrows: Parakeet gets 8.04 percent ³⁵ against Apple's 10.34 percent.

What the data cannot tell you

Several limits apply, and they point in different directions.

• Style is not error. FLEURS sentences follow the punctuation of their source, and a comma can be optional. Part of every formatted figure is a style difference, not a mistake. This affects every engine alike, so comparisons hold better than absolute levels.
• Digits and words. Numbers written as digits against words count as errors. A dictation app that writes "twenty" where the reference has "20" is penalized, though a reader may not care.
• Read speech, not your speech. FLEURS is read aloud in clean conditions, 100 recordings per language. It says little about accents, noise, interruptions or the way people actually compose messages.
• One run, one model version each. These are single measurements, with no spread or confidence interval. Differences of a point or less, such as German at 7.06 percent for Parakeet against 7.10 percent for Whisper, should be read as ties.
• A model is not an app. Apps add their own punctuation rules, voice commands and post-processing. A figure here describes the raw engine output as we scored it.
• Unequal coverage. Apple has no row where it does not support the language, so its median cannot be compared with the other two over the same set of languages.

What this means if you dictate

Start from your language, not from the engine. In Spanish, Italian, German, Russian and Portuguese, expect light edits to punctuation and capitals, and any of the three engines is workable where it is supported. In the Baltic languages, Slovenian and Hungarian, expect to rewrite a large share of the text, and check Parakeet first. For Greek, Whisper is the engine to try. For Maltese, avoid Whisper.

If you work on a Mac in one of Apple's languages, the built-in engine gives steady results and is a fair default. Check Portuguese separately before relying on it. Treat dictation as a first draft in every language, and plan a pass for commas and capitals. Our overview of speech recognition explains why those marks are the hardest part of the output to get exactly right, and the on-device benchmark shows how the same engines do when only words are counted.

Showing 8 of 25 rows. See every row in the dataset.