Read the marketing for paid dictation on a Mac and a pattern shows up quickly: most of these apps do not promise to type what you said. They promise to type something better than what you said. Filler words removed, punctuation inserted, grammar corrected, the register adjusted for where the text is going, a rambling thought reshaped into a structured one. As of 10 September 2026, Wispr Flow advertises automatic removal of filler words, added punctuation and formatting, plus style settings; Aqua Voice advertises real-time refinement of phrasing and formatting chosen to suit the destination; superwhisper ships modes that optimise tone, structure and formatting; MacWhisper lists automatic cleanup and grammar improvement as Pro features. Five of the six paid tools in this category treat rewriting as the product.
That is a defensible position and for a lot of people it is the right purchase. This page is the argument for the other one, and — more usefully — a way to find out which kind you are currently running, because the difference is invisible until it costs you something.
The failure nobody warns you about
A transcription error is easy to notice. You said "Postgres" and the screen says "post grass", and you fix it, and you are mildly annoyed. The error is visible because the output is obviously wrong.
An edit is not like that. You said:
I think we probably want to hold off on the migration until the audit's done, but I could be talked out of it.
and what lands is:
We should delay the migration until the audit is complete.
Nothing is misspelled. Nothing is ungrammatical. It reads better than what you said. And it means something different: two hedges are gone, and a sentence that invited an argument now closes one. If you send that to your team, they will act on a position you did not take.
This is worth separating from accuracy, because it is not an accuracy problem and no amount of a better engine fixes it. The words were heard correctly. They were then improved.
I think we probably want to hold off on the migration until the audit's done, but I could be talked out of it.
When your phrasing is the content
There is a class of writing where the exact words carry the work, and an editor that runs before you see the text is a liability rather than a feature.
Anything with a voice. A novelist's dialogue, a newsletter that sounds like a person, a screenplay — the whole value is the specific way it is put. A model that normalises phrasing toward well-formed prose is deleting the asset. If you write anything where somebody would notice you had been replaced, see dictation for writers.
Anything where hedges are load-bearing. "Probably", "as far as I can tell", "one option would be" are not filler. In a clinical note, a legal memo, a code review or a status update, the difference between "this will work" and "this should work" is the difference between a claim and an estimate.
Prompts. This one surprises people until it happens to them. When you dictate an instruction to an AI assistant, your phrasing is the input, and a tidying pass rewrites the thing you were trying to send. "Don't refactor anything else" and "avoid unnecessary changes" produce different results. Dictating prompts is the use case where a helpful editor does the most damage and is hardest to spot, because the output you are judging is two steps downstream.
Quotes and verbatim capture. A journalist reading back a quote, a researcher transcribing an interview note, a support agent recording exactly what a customer said. Verbatim means verbatim.
Bilingual writing. If you mix languages in a sentence on purpose, a cleanup pass frequently decides one of them was a mistake.
When polishing is the better buy
A comparison that only lists the other side's weaknesses is an advertisement, so here is the honest half.
If you dictate a hundred emails a week to people you do not know well, an app that reliably produces clean, formatted, professional paragraphs is saving you a real edit pass on every one. If you write in a second language, a tool that fixes agreement and article errors as you go is removing a tax you pay all day. If you think out loud and want the structure imposed afterwards, that is exactly what those modes are for. And if your dictation is mostly short messages where tone matters more than precision, "make this sound casual" is a genuinely useful control.
Some of those tools are also better than this one at things unrelated to rewriting — four platforms, a vocabulary that learns, a hundred languages. The full comparison against Wispr Flow works through where each of those matters, and the one against Aqua Voice does the same for real-time refinement.
The argument here is narrower than "polishing is bad". It is that polishing should be a thing you chose, and in most of these products it is the default you never consciously agreed to.
How to find out what your current app does
Four minutes, and it works on anything.
1. Dictate a sentence with two hedges and a self-correction. Say: "I think we should probably ship on Thursday — actually, Friday — assuming the tests pass." Then read what landed. If the em-dash correction became a clean "Friday" with no trace of Thursday, the tool is editing. If "I think" and "probably" survived, it is not.
2. Say a filler word deliberately. "Um, so the thing is, uh, we already tried that." Count the ums in the output.
3. Dictate one long unpunctuated run-on. Speak sixty words without pausing and without saying any punctuation. A tool that returns tidy sentences with commas and full stops made structural decisions for you. Whether you want that is the question; that it happened is the finding.
4. Say something slightly awkward but correct. "The thing what he said was wrong" — deliberate bad grammar. If it comes back correct, grammar correction is on, and you now know it can also "correct" a technical term or a name it did not recognise.
5. Check whether the original still exists. This is the one people skip and it matters most. If the app processes your speech, does it keep what it heard before processing? If not, an edit you disagree with is unrecoverable — you cannot compare against something that was never stored.
What this app does instead
Coii VoiceInput's post-processing is a fixed chain, not a model. In order: run together repeats from the decoder are collapsed; your own dictionary is applied; filler words are removed if and only if you turned that on; and punctuation and spacing conventions are applied last, because they depend on the final word sequence. That is the entire list. Nothing in it takes a phrase and produces a different phrase with the same meaning, because nothing in it knows what a phrase means.
Filler removal ships off. That is a deliberate default rather than an oversight: people who dictate verbatim have no way to know their words are being deleted, and silently deleting what somebody said is not a thing to opt people into.
Every history row stores both the raw transcript and the processed one. If a dictionary entry fires where you did not want it, the original is one click away rather than gone. That is what makes it safe to have any filters at all.
The dictionary is the part you control: names, product terms and rewrite rules
like btw → by the way, typed in by you. It is used twice — as a hint to the
engine before recognition, and over the transcript afterwards — and it does
exactly what you entered, on word boundaries, never fuzzily. It does not
learn. If you have fifty internal terms and no patience for entering them, a
tool that picks them up as you go is genuinely the better buy, and that row is
marked as a loss on every comparison page here.
Predictable beats clever
There is a second reason to prefer a fixed chain that has nothing to do with voice or style, and it is the one engineers tend to land on first: you can learn it.
A chain of four filters does the same thing on Tuesday that it did on Monday. When something comes out wrong you can work out which stage did it, and you can switch that stage off. A model that decides how to present your sentence is making a judgement each time, and the same input can produce a different result next month without anything on your machine having changed. That is fine for a draft and awkward for a tool you use four hundred times a day, because the mistakes are not reproducible and therefore not fixable — you cannot file a bug against a judgement call.
It also means the failure modes here are boring. The worst thing a punctuation filter does is put a full stop where you wanted a question mark. The worst thing an editing pass does is change what you meant, and you will not necessarily notice.
The specific case of names it has never heard
A general engine gets unusual surnames, internal product names and technical terms wrong. Every engine does, this one included, and it is the single most common accuracy complaint about every dictation product ever sold. The question is what the app does about it.
An app with a vocabulary that learns watches what you correct and picks the terms up over time, which is genuinely less work. This one gives you a dictionary to type them into instead: the term, and optionally what it should become. Entries are used as a decoding hint before recognition — so the engine is biased toward hearing the word correctly in the first place, which is better than fixing it afterwards — and applied again over the transcript. Latin-script entries match on word boundaries and case-insensitively; nothing is fuzzy or phonetic, because an entry that rewrote things that merely sounded similar would corrupt sentences that had been transcribed correctly, and you would have no way to tell which ones.
Teaching it your jargon walks through building that list for a real stack, and is worth twenty minutes if your work has vocabulary in it.
The rest of the shape
This app runs entirely on your Mac. Nothing is uploaded — no audio, no transcripts — and it works with the network off once the weights have been fetched the first time. It is $19 once for macOS 13 or later, with a 30-day trial and no account. The related argument about what a cloud tool does with your voice is worked through in the apps that don't train on your voice.
Two Mac tools besides this one leave your words alone. VoiceInk is open source and a one-time purchase, though its own feature list includes filler and repetition removal and restructuring, so it is in the editing camp by default. Handy is free, open source and does nothing but transcribe and paste — if the only thing you want is verbatim text and no cost, look at it before you look at this.
Who should pick which
Pick a tool that rewrites if most of what you dictate is correspondence, you want formatting decided for you, you write in a second language, or you would rather edit a clean draft than a faithful one.
Pick one that doesn't if you write anything with a voice, dictate prompts, work where a hedge changes the meaning, capture quotes, or simply want to know that the text on screen is the text you said. Try it for thirty days — the honest test is whether you ever miss the tidying.
Whichever way that goes, run the four-minute check above on what you are using now. Most people have never actually looked, and the finding is usually a surprise. If you are new to the category, what dictation software is covers the ground underneath this argument, and the MacWhisper comparison is the closest thing to a like-for-like on price.