Speak a sentence out loud and you will not say a single comma. You pause, you breathe, your pitch falls at the end of a thought — but the punctuation lives entirely in the writing, not the speech. That gap is the oldest problem in dictation, and for a long time the answer was to make you close it yourself: you learned to say "comma," "period," "new paragraph" out loud, threading typography through your own sentences like stage directions. It worked until it didn't — and when it didn't, the word comma landed in your text as the literal word. Modern dictation takes the opposite bet: that a system which understands the sentence can put the marks in for you.
Saying the punctuation out loud was the tax
The old model treated punctuation as dictation you performed. Every clause boundary meant interrupting the thought you were capturing to issue a command — "the deck is ready comma send it by five period," the sort of line you'd hit constantly when dictating email. You cannot compose a sentence and narrate its punctuation at once without splitting your attention, and dictation is supposed to keep your focus on the idea. The moment you're managing commas by voice, you're doing two jobs at once.
And the failure mode was worse than silence. When the system misheard the command, the word appeared literally: you meant to end a line and got the text "new paragraph" sitting in the middle of your sentence. So you stopped, deleted it, and lost the thread you were mid-way through. A tool that punishes you for using it exactly as instructed teaches you to stop trusting it.
What "inferring" punctuation actually means
Modern dictation runs a pass over the whole utterance instead of committing word by word — the second-stage rewrite in how AI dictation works under the hood — and punctuation falls out of that pass the way it falls out of reading — from structure and meaning, not from spoken commands. Prosody carries real signal: the short pause that separates two clauses, the longer settling that ends a thought. But rhythm alone is unreliable — people pause to think mid-sentence and rush straight through real boundaries — so the useful systems weight the words too. "Send the deck by five and let me know" is one sentence; "send the deck by five. Let me know when it lands" is two, and what decides is the grammar, not just the gap between the words.
The practical result is that you stop performing typography and just talk. A question comes back with a question mark because it was phrased as one. You are no longer dictating punctuation; you are dictating meaning, and the marks are inferred from it.
Paragraphs and lists are formatting, not punctuation
The same inference reaches past the comma. When you talk through three options in a row — "first we cut scope, second we push the date, third we add a person" — that has the shape of a list, and a system reading for structure can return it as one instead of a wall of clauses. When you finish a thought and move to a new one, that's a paragraph break, not a comma. This is the difference between a transcript, which is what you said, and a draft, which is something you could send — and the gap between them is exactly this formatting layer, which is why it matters so much when you're drafting long-form by voice.
The discipline is knowing when not to format. Not every pause is a paragraph, and a tool that over-structures your speech — chopping a single flowing thought into staccato fragments — is as wrong as one that returns an undifferentiated block. Good inference is quiet: you notice it only by its absence.
East Asian punctuation is a different, harder problem
Here the naive approach doesn't just underperform — it produces text a native writer would never send. East Asian scripts use full-width punctuation that is visually distinct from the Latin marks: the full stop is 。 not a period, and the exclamation and question marks are ！ and ？, sized to sit among Han characters rather than crammed against them. A system that drops a half-width Latin period into a line of Chinese or Japanese leaves a mark that looks wrong to anyone reading it — small, but unmistakably foreign.
The comma is worse, because Chinese and Japanese have two of them and they are not interchangeable. The clause comma ，separates parts of a sentence, where an English comma would go. But a list of nouns takes the ideographic comma 、(the "enumeration comma") instead — 蘋果、橘子、香蕉 for "apples, oranges, bananas." Nothing in the sound tells you which one a boundary needs; it's a grammatical choice, resolved from whether the items are clauses or a list, which means only a system reading structure can get it right. A tool that pours every pause into one Latin comma has quietly marked its output as machine-made to a fluent reader.
The mixed-script line is where it all comes due
Real writing in Chinese, Japanese, or Korean rarely stays in one script. English product names, acronyms, and numbers sit inside the sentence, and each switch is a decision about which punctuation system governs the space around it — the spacing where full-width meets half-width, the mark after a Latin acronym mid-clause. Get these wrong and the line reads as stitched-together rather than written. This is the compounding case: it needs the list inference and the full-width-versus-half-width judgment and an understanding of which language owns each span, all at once, from speech that never announced any of it.
The test is whether you can just talk and read it back
None of this shows up in a feature list, and none of it is something you should have to manage by voice. The honest test is the plain one: talk the way you actually talk — your real pace, your real pauses, your real mix of languages — and read back what comes out. Are the commas where you'd have put them? Did the list come back as a list? In a CJK line, is that 、and not a Latin comma, the full stop 。and not a period? If you can dictate a thought and read back something you'd send without touching the punctuation, the formatting layer is doing its job. If you're still fixing marks after the fact, you never stopped typing — you just moved it to the end.