Formatting passpost-processing
The formatting pass is the step after speech recognition that turns a raw transcript into text a person would have typed: punctuation, capitalization, numbers as digits, fillers removed, and a tone suited to where the text is going. It is where most of the difference between dictation tools lives.
A recognizer produces words. It does not know that "ten thirty am" should be 10:30 AM, that a Slack message is lowercase and a client email is not, or that "bob at acme dot com" is an address. The formatting pass knows, or is supposed to. Early tools did it with rules; the current generation runs a small language model over the transcript; the best do both, with rules for anything that has one right answer and the model for judgment.
It is also the step with the most ways to go wrong. A model asked to "clean this up" will reword, tighten, and occasionally invent, and a slow model turns a fast tool into a waiting room. How a product constrains its formatting pass tells you more about its quality than which recognizer it uses.
What a good pass does
Removes fillers and false starts. Punctuates and capitalizes. Writes numbers, times, dates, prices, units, emails, and URLs in their proper form. Applies spoken commands like "new paragraph". Inserts snippets and dictionary terms exactly as saved. Matches the tone of the destination app, casual or professional, and can be pinned per app. Leaves terminals and code editors untouched. Everything on that list is either a rule or a narrowly briefed model, and the order matters: rules first, so the model never gets to second-guess a clock time.
Rules versus the model
A language model is good at judgment and unreliable at precision. Given "ten thirty AM" it usually writes 10:30 AM, and "usually" is the problem, because you will not notice the time it wrote 10:30 PM until the meeting is missed. Anything with one correct answer, clock times, spoken punctuation, dictionary words, snippets, symbol names, should be handled by a deterministic rule that produces the same output every time and can be tested. The model gets what is left: sentence flow, paragraph breaks, which fillers were filler, and tone.
Guardrails
Three are worth insisting on. A content check: if the model’s output loses a word you said or adds one you did not, discard it and keep the rule-formatted transcript. A time budget: if the model has not answered within a fixed window, the transcript lands without it. And a verbatim escape: apps where any rewrite is a bug get no model at all. With those in place, a formatting pass can only improve the text or leave it alone; it can never break a dictation.
In Flit
Flit’s pass runs on your Mac: fixed rules for everything exact, a local model for judgment, a content check on its output, and a time budget so a late model never holds the text.
The details are not left to a model →Questions
Fair questions.
What is post-processing in dictation software?
Everything that happens to the transcript after the speech model produces it: punctuation, capitalization, number formatting, filler removal, spoken commands, and tone. It is the stage that turns a transcript into writing, and the stage where tools differ most.
Does the formatting use AI?
In most current tools, partly. A small language model handles judgment calls like tone and sentence flow. The better tools handle anything with one right answer, such as times and punctuation, by fixed rule instead, because a model is only usually right.
Can I turn formatting off?
A good tool lets you choose the level, from verbatim to fully polished, and switches to verbatim automatically in terminals and code editors. Less formatting also means faster text, which is a reasonable trade for a quick reply.
Where this shows up on flit.fyi
Related terms
