Why Raw Dictation Needs a Second Pass

Here is a sentence as it actually comes out of a person's mouth: "so um I wanted to let you know that the the design review moved to Thursday at three and uh can you send the updated mockups to Sarah before then".

Every word of that is correctly transcribed. It is also not something you would send to anyone. The transcript is accurate and unusable at the same time, and that gap is the whole argument for a second pass.

Speech is not writing, and transcription does not convert between them

A transcription model has one job: turn the sounds into the right words. It does that job well, and doing it well produces an artifact shaped like speech, because speech is what it heard.

Spoken language has structure that written language does not. You hold the floor with fillers while you think. You start a sentence, abandon it, and start again. You correct yourself mid-clause and expect the listener to keep the second version. You signal ends of thoughts with pitch and pause rather than periods. All of that is normal and fluent in conversation, and all of it is noise on a page.

So a raw transcript reliably contains:

• Fillers: um, uh, so, like, you know, right, used as timing rather than meaning.
• False starts: "I wanted to, can you send," where the first clause was abandoned and never finished.
• Self-corrections: "Thursday, sorry, Friday at three," where only the second version is the real content.
• Repeated words: "the the," "that that," an artifact of thinking and talking at once.
• Little or no punctuation: no sentence boundaries, no paragraph breaks, and often no capitalization to speak of.
• Spoken formatting commands: "new paragraph," "period," "bullet point," sitting in the text as literal words.

None of that is a transcription error. It is a faithful record of how people talk.

What the second pass should change

The cleanup step in Voice is scoped deliberately narrowly. It removes filler words, stutters, and self-corrections, fixes punctuation, and follows spoken commands like "new paragraph." That is mechanics. It is the pass a copy editor would make without asking you any questions.

The test for whether a change belongs in that pass is simple: would you have made this exact change yourself, without thinking, in under a second? Deleting an "um" passes. Adding a period where you clearly stopped passes. Turning your spoken "new paragraph" into an actual break passes. Rewriting your sentence because it could be tighter does not.

What it must leave alone

This is the more important half, and it is where most tools in this category overreach.

Your words should survive. If you said "get," the cleanup should not give you back "obtain." If you said "we should probably push the deadline," it should not give you back "we will extend the deadline." Hedges, qualifiers, and the specific degree of commitment in a sentence are content, not clutter. Someone dictating a note about a delicate situation chose "probably" on purpose.

Your tone should survive. Dictation is disproportionately used for messages to people who know you. A cleanup that levels every message into neutral business prose is actively making your writing worse, and it does it invisibly, because the result reads fine in isolation. You only notice when you realize you have stopped sounding like yourself.

And nothing should be added. A model asked to improve text will sometimes supply a detail that fits: a plausible time, a connecting reason, a smoothing clause. Every one of those is a sentence you now have to fact-check, in a document you were about to send. A pass that only removes and punctuates gives you nothing to verify.

Why you have to be able to see the original

There are two places dictation can go wrong, and they stack.

First, the transcription can mishear. Second, the cleanup can misjudge which words were filler and which were meaning. The second failure is subtler than the first, because a mistaken cleanup produces fluent text. A misheard word often looks strange on the page and catches your eye. A dropped qualifier does not look like anything at all.

That is why the Voice panel shows the cleaned version and the original together, with a toggle between them, before anything reaches your clipboard. You read the clean version because it is faster to read. You check the original when a sentence feels slightly off, or when the content is sensitive enough that "slightly off" is not acceptable. The word count sits at the top, and Discard and Redo sit next to Copy, so backing out is as easy as accepting.

An interface that hides the original is asking you to trust a rewrite you cannot inspect. Even when the rewrite is good, that is the wrong habit to build.

When to skip the pass entirely

Cleanup is optional in Voice for a few reasons, and they are worth knowing so you use the switch.

Skip it when the exact words matter more than the readability: quotes you intend to attribute, a statement you are recording for a file, anything where a removed hedge changes the meaning of the record. Skip it when you are dictating fragments that are supposed to be fragments, like a list of bullet points or a set of search terms, because a cleanup pass will try to make them into sentences. And skip it for anything private enough that you would rather the text not go anywhere, since cleanup is the step where the transcript is sent to Claude with your own API key, while transcription itself happens on your Mac.

The raw transcript is not a degraded product. For a lot of uses it is the correct one.

The takeaway

Treat the second pass as punctuation and deletion, not as editing. The moment a cleanup step starts improving your sentences, you have traded a small chore, adding periods, for a large one, verifying that the text still says what you said. The tools worth using in this category are the ones with a narrow definition of their job and a visible original sitting one click away.