Does dictation get punctuation and capital letters right?

Dictation gets punctuation and capital letters mostly right, but not well enough to skip proofreading, and how much you fix depends on the language more than on the engine. Across our October 2026 test, the median formatted error rate was 12.75 percent for Parakeet TDT v3 ¹, 12.94 percent for Whisper large-v3-turbo ² and 10.62 percent for Apple SpeechTranscriber ³. Lower is better, and each figure counts errors per 100 tokens.

The formatted error rate counts punctuation marks and capital letters as a reader sees them. The word error rate (WER) counts only the words. Putting the two side by side shows how much work the formatting adds.

Counting punctuation and capitals always raises the error rate

In English, Whisper large-v3-turbo scored 7.71 percent formatted ⁴. Its words-only score, in the "Words only (WER)" column of the table below, is lower. Parakeet TDT v3 scored 8.18 percent ⁵ and Apple SpeechTranscriber 11.10 percent ⁶.

The pattern holds in the other large languages. Spanish was the cleanest: Parakeet TDT v3 reached 5.44 percent ⁷ and Whisper 5.68 percent ⁸. Italian was close behind for Parakeet at 6.40 percent ⁹, and German sat near 7.06 percent for Parakeet ¹⁰ and 7.10 percent for Whisper ¹¹.

The gap between formatted and words-only scores is large in big languages and small in hard ones

In Spanish, Parakeet's formatted score was 5.44 percent, while the words alone were wrong far less often, which is why the "Words only (WER)" column matters. The words were nearly right, and the punctuation and capitals made up a large share of what remained.

In harder languages the picture flips. Whisper's Estonian text scored 20.52 percent formatted ¹², and the words-only figure sits just below that in the data. There, wrong words dominate and formatting adds little. Latvian and Lithuanian look the same: Whisper reached 20.35 percent ¹³ and 26.71 percent ¹⁴.

Language matters more than engine

Parakeet and Whisper finish within a fraction of a point of each other at the median, but their spread across languages is wide. Parakeet's figures run from 5.44 percent in Spanish to 34.83 percent in Greek ¹⁵. Whisper's run from 5.68 percent in Spanish to 71.96 percent in Maltese ¹⁶.

The Maltese row is a failure of the words, not of the formatting: the words-only score was higher still, at 71.96 percent formatted against a plain figure shown in the table. Treat Whisper as unusable for Maltese in this test.

Here are the Whisper and Parakeet rows for a few European languages:

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

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

Apple's model covers only some languages, and one row failed outright

Apple's model supports six of the 25 languages, so its median of 10.62 percent rests on just six rows. Its best figure was French at 10.34 percent ²⁶, followed by Spanish at 10.43 percent ²⁷ and German at 10.55 percent ²⁸.

The Portuguese row reads 100.00 percent ²⁹, and its words-only score is also the maximum. That means the output did not match the reference at all in our run. The data does not say why, so do not read it as a statement about Apple's punctuation. It does explain the wide range behind Apple's median.

What these numbers do and do not measure

The benchmark uses FLEURS, a set of read sentences, scored once all 100 recordings of a language were done. Each row compares an engine's text with the written reference sentence. Three limits apply:

• Style counts as error. FLEURS sentences follow their source's punctuation, and a comma can be optional. Part of every gap is style, not a mistake, and it hits every engine alike.
• Digits against words count. If the reference writes a number as words and the engine writes digits, that is an error, even though many readers prefer digits.
• Read speech is easier than real speech. These are clean, read sentences. Spontaneous dictation, background noise and spoken commands like "new paragraph" are not tested.

So the figures compare engines and languages with each other. They are not a promise of how many edits your own text needs. The dataset holds 56 data points ³¹, and you can inspect each one in the full dataset.

What this means for choosing how to dictate

• Dictating in English, Spanish, Italian, German, French or Portuguese: any of the three engines is workable for punctuation. Plan a proofreading pass, not a rewrite.
• Dictating in a Baltic, Slavic or Finno-Ugric language: the words are the bigger problem. Test with your own voice before committing, and compare engines per language rather than trusting a headline median.
• Choosing a tool: check whether it uses a model that covers your language, then compare products in our market overview. Apple's built-in option only covers six of these languages.

For the raw word accuracy behind this, see the on-device ASR benchmark. For background on how recognition error is measured, read the guide to speech recognition accuracy and the overview of speech recognition.