← Back to Blog

Short answer: Edit dictated text in a separate pass from dictating it. Speak the whole draft without stopping, then fix it: use Voice Edit to speak a correction instead of placing the cursor, add replacement rules for words the engine gets wrong every time, and re-dictate a whole sentence rather than repairing it word by word.

Most people who quit dictation do not quit because it was inaccurate. They quit because of what happened after. The words appeared fast, and then they spent four minutes dragging a cursor around trying to fix three of them, and concluded the whole thing was slower than typing.

That conclusion is understandable and it is wrong, but only if you change how you edit. Dictation and typing produce text at different speeds, and they also produce different kinds of mistakes. If you edit dictated text the way you edit typed text, you will be slow. The fix is not a better microphone or a better engine. It is a different workflow.

This guide covers what actually goes wrong with dictated text, the four repair techniques worth knowing, and how to set things up so the same error stops coming back.

Why editing dictated text feels harder than it should

When you type, you edit continuously. Your hands are already on the keys, your eyes are already on the line you just wrote, and correcting a typo costs you one backspace and half a second. Editing is so cheap that it disappears into the act of writing. You are not aware of doing it.

Dictation breaks that loop in three ways.

Your hands are not in position. The whole point was to keep them off the keyboard, or to be walking around, or to have a phone in one hand. The moment you need to fix something, you have to re-enter the typing posture you were trying to avoid. On a phone, that means the cursor-placement problem: trying to tap between two specific letters with a fingertip that covers six of them.

The errors are semantic, not mechanical. A typo looks wrong. teh jumps off the page. A dictation error is usually a real, correctly spelled word in the wrong place. "Their" for "there." A client's surname turned into a common noun. A number heard as a homophone. Your eye slides right over these because nothing is visually broken, which means proofreading dictated text takes more attention per line than proofreading typed text.

The volume is higher. A comfortable adult types around 40 words per minute; professional typists reach 80 to 100. Ordinary speech runs 130 to 150 words per minute. So a five-minute dictation session hands you a wall of text that would have taken fifteen or twenty minutes to type. Even at high accuracy, more words means more absolute errors sitting in front of you at once. That is not a failure of the engine. It is arithmetic, and it feels like failure if you were expecting a finished document.

Dictation does not give you a finished document. It gives you a fast first draft, which is a different and more useful thing.

Rule one: never edit while you dictate

This is the single change that fixes most of the problem, and almost everyone resists it at first.

The instinct is to fix errors as they appear. You see a wrong word land on screen, you stop, you correct it, you resume. This feels efficient. It is the most expensive thing you can do, for two reasons.

First, it destroys the thing dictation is actually good at. Speaking is fast because it is continuous. Every stop costs you the sentence you were building in your head, and you spend the next ten seconds rebuilding it. Ten stops in a five-minute session and you have lost more time to context switching than the errors would ever have cost.

Second, you fix things that would not have survived anyway. Roughly a third of the sentences in any first draft get cut or rewritten. Polishing a sentence you are about to delete is pure waste.

So: speak the entire draft. Watch the screen if you like, but do not touch anything. If you notice a bad error, keep going. If you notice you said the wrong thing entirely, just say the sentence again correctly and keep moving, because deleting the duplicate later takes three seconds and stopping now costs thirty.

Then, when you are done talking, switch modes deliberately. Read the whole thing top to bottom, then repair. Two passes, cleanly separated. This is how professional dictation has worked for a century, and it is why the transcriptionist role existed at all.

Rule two: speak the correction instead of hunting for it

The most painful part of editing dictated text is not deciding what to change. It is navigating to the place where the change goes. On a phone this is genuinely awful: tap, magnify, miss, tap again, accidentally select a word, tap again.

Voice Edit exists specifically to remove that step. Instead of placing a cursor, you select the text and speak the change you want in plain language. "Change the second sentence to past tense." "Replace 'quarterly' with 'annual'." "Make this shorter and less formal." The text is rewritten in place. You never touched the cursor.

This inverts the usual cost model. In normal editing, small local fixes are cheap and big structural rewrites are expensive. With spoken editing it is closer to flat: rewriting a paragraph costs about the same as fixing one word, because the effort is in describing the change, not in executing it. That has a practical consequence worth internalizing.

Prefer the bigger fix. If a sentence has two errors in it, do not fix the two errors. Re-dictate the sentence. If a paragraph is muddled, do not repair the seams. Say the paragraph again. The version you speak the second time is usually better than the patched original, because you now know what the paragraph is for. Patching preserves the structure of a draft you wrote before you knew where it was going.

What spoken editing is good at

What it is not good at

Rule three: fix recurring errors at the source

Here is the distinction that separates people who enjoy dictation from people who tolerate it.

There are two kinds of dictation error. Random errors happen once, usually because you mumbled, or a door slammed, or you trailed off at the end of a sentence. You fix these in the edit pass and forget them. Systematic errors happen every single time, and they are almost always the same category of word: proper nouns, brand names, internal jargon, technical terms, and abbreviations that only your industry uses.

Systematic errors are the ones that make people give up, because fixing the same surname on the same client's file every day for a month is demoralizing in a way that random typos are not. And they are the ones you should never fix by hand more than twice.

Smart Vocabulary on Mac is the tool for this. It is a personal dictionary with replacement rules: you tell it the wrong output and the right output, and the correction happens automatically from then on. Add your colleagues' names. Add your product names. Add the three-letter acronyms your team uses. Add the specific mangling the engine produces for each, since the point of a replacement rule is that it catches the exact wrong version you keep getting.

A useful habit: keep a scratch note open for your first week. Every time you fix something during an edit pass, ask "will this happen again?" If yes, it goes in the vocabulary list, not just in the document. Ten minutes of setup in week one removes most of your editing burden for the rest of the year. We wrote more about this in how custom vocabulary learns your words.

Hardware and environment cover a related slice of the systematic problem. If every session has errors clustered at the start of sentences, you are probably starting to speak before the recording begins. If the errors are spread evenly and get worse in the afternoon, it may be the room. Fixing background noise and improving accuracy with an accent both remove whole classes of error before they reach the page.

Rule four: dictate the structure, not just the words

A surprising share of "editing" is not correcting mistakes at all. It is formatting. You dictated a wall of text and now you are inserting paragraph breaks, turning three clauses into bullet points, and adding the headings that were implied by your talking but never made it to the page.

This is entirely avoidable. Say the structure out loud as you go. Paragraph breaks, list items, and new lines are all spoken commands, and using them costs you nothing during the draft pass while saving several minutes of layout work afterwards. The mechanics are covered in dictating new lines and paragraphs and dictating bullet points and lists.

Punctuation follows the same logic, with one caveat. Speaking punctuation feels unnatural for about a week and then becomes invisible; almost everyone who sticks with dictation ends up doing it, because the alternative is inserting commas by hand into eight hundred words. If you are getting the literal word "period" in your text instead of a full stop, that is a different problem with its own fix, and it is worth solving before you blame the engine for anything else.

Capitalization is the third member of this family. If proper nouns come back lowercase, that is usually not a transcription failure but a formatting one, and it is fixable at the settings level rather than one word at a time. See Mac dictation not capitalizing for the specific causes.

A practical two-pass workflow

Putting the rules together, here is a workflow that holds up for anything longer than a text message.

Pass one: produce

  1. Decide the shape before you speak. Even ten seconds of "intro, three points, ask" beats going in cold, because rambling is the single biggest source of editing work.
  2. Hold the hotkey and talk until you are finished. On Mac, hold, speak, release, and the text lands at your cursor in whatever app you are in. Do not stop for errors.
  3. Say the structural commands as you go: paragraph, new line, bullet.
  4. If you go badly wrong mid-sentence, say the sentence again correctly and keep moving. Leave the wreckage for pass two.

Pass two: repair

  1. Read the whole thing once, top to bottom, without fixing anything. You are building a map of what needs to change, and you will make better decisions than you would fixing in order.
  2. Delete first. Cut the duplicate sentences, the false starts, the paragraph that repeats the one above it. This is usually the largest single improvement and it requires no correction at all.
  3. Fix structural problems next. Reorder, merge, split. Speak these as changes rather than executing them by hand where you can.
  4. Then and only then do the word-level pass: names, numbers, homophones.
  5. Log anything that will recur into your vocabulary list.

Doing deletion before correction matters more than it sounds. If you correct first, you will carefully repair sentences that you are about to delete in step two. Cut, then fix what survives.

The errors worth checking every time

After the structural pass, dictated text tends to fail in a small number of predictable places. A checklist beats a general "read it carefully," because your eye genuinely does not catch these on its own.

That last one is worth a habit rather than a check: hold the key for a beat after you finish speaking. It costs nothing and removes a recurring error.

Editing on a phone specifically

Everything above applies on iPhone, but the cost of manual editing is several times higher, so the rules bind harder.

The keyboard's mic button lets you dictate into any app, which means the draft pass is easy. The trap is the repair pass: precise cursor placement on a touch screen is the worst text-editing experience on any device most people own. This is where speaking the correction stops being a preference and becomes the only sane option. Select the sentence, say what should change, move on.

The other phone-specific habit is knowing when to stop. Phones are for capture. If a message needs three rounds of careful revision, capture it on the phone, finish it on a laptop. Trying to do fine editing with a thumb is how people end up sending the version with the wrong client's name in it.

When translation is part of the edit

If you write to people in other languages, the edit pass usually includes a translation step, and doing it as a separate step doubles the work. Dictating with two-way translation turned on collapses the two: you speak in the language you think in, and the text arrives in the language you are writing in, across 24 languages. What you then edit is a draft in the target language rather than a source text plus a translation task.

The editing checklist above still applies, with proper nouns deserving extra attention since names and product terms are exactly what survives translation least reliably. Replacement rules help here too.

How much editing is normal

A reasonable expectation, once your vocabulary list is populated and you have stopped editing mid-flow: a short email needs a glance and one or two touches. A long document needs a real read-through, the same as anything you typed. The ratio that matters is not errors per hundred words, it is total minutes from "I need to write this" to "sent," and that is where dictation wins even when the raw text needs work.

If you are spending more time editing than you would have spent typing the whole thing, something specific is wrong, and it is nearly always one of four things: you are editing mid-flow, you are not using replacement rules for recurring terms, your environment is adding errors, or you are dictating without deciding what you want to say first. All four are fixable, and none of them are about the engine.

Getting set up

Voice Keyboard Pro works the same way on both platforms and is built around the two-pass approach described here.

On Mac it lives in the menu bar. Hold your hotkey, speak, release, and the text appears at your cursor in any application, so there is nothing to enable per app. Smart Vocabulary holds your personal dictionary and replacement rules, and Meeting Mode handles the longer sessions where speaker detection and notes matter. On iPhone it is a keyboard with a mic button, available anywhere you can type, with Voice Edit for spoken corrections and two-way translation for 24 languages.

On privacy, since dictated drafts are frequently the most sensitive text a person writes: our server stores only operational pings. No audio and no transcript content.

There is a free tier with daily limits, and Pro is $4.99 a month or $34.99 a year. The most useful first exercise is not a test document. Take something real that you were going to type this afternoon, dictate it straight through without stopping once, and then time the repair pass. That number, not the raw accuracy, is the one that tells you whether this works for you.

Keep reading