Writing

Without any AI, MicJot writes numbers, money, times, percentages, email addresses, punctuation and line breaks the way you would type them. When it cannot be sure, it leaves your words exactly as you said them.

The rule

A number or time left in words is never lost text; a wrong conversion is. So every rule below converts only when the sentence gives it a cue, and otherwise keeps the spoken words. These examples are from the formatter's own tests and release notes:

What the rule-based path does

• Numbers and money: digits, thousands separators, currencies (“four hundred and twenty rupees fifty paise” → ₹420.50), Indian scales (“two crore forty lakh” → 2,40,00,000), ranges that share a scale word only when the reading ascends (“two or three hundred people” → 200 or 300 people), decimals that convert whole or not at all.
• Times and dates: a time converts when something marks it as one: a meridiem, an event word (meet, call, flight, standup…), a weekday or “tomorrow”, or the end of the sentence. Ordinals and months convert in dates.
• Equations, percentages, degrees, phone numbers, email addresses and URLs are written as typed; “over” is a division sign only inside an equation.
• Spoken punctuation and layout: period, comma, colon, question mark, “new line”, “new paragraph” and “bullet point”; the noun “period” after a determiner stays a word (“a tough period”).
• Cleanup: filler words (“um”, “uh”) and stutters are removed and sentence capitals fixed; in languages other than English, or when the language is detected as one, formatting and filler removal are skipped so a German “um” stays.
• Styles: a tone per app category (personal messages, work messages, email, everything else), applied by the rule-based path (capitalisation, final punctuation) and, with AI writing, by the model. Any app can be reassigned.
• Dictionary and snippets: your words reach the recogniser as a prompt (up to 64 tokens per 30-second window, starred entries first) and replace misheard forms; a snippet phrase expands into its text.

Known limitations

The complete list ships with each release; these are the ones you are most likely to meet in 0.3.1.

• Times without a cue stay in words. “I have booked a table at seven thirty” keeps the words; add “pm” or an event word and it converts. A time dictated right after text you typed does not see the typed event word.
• Levels and scores stay words. “we are at two fifty” keeps the words unless “am”, “pm”, “sharp” or “latest” follows.
• Shorthand prices stay words. “seven ninety nine” stays as spoken; “seven hundred ninety nine” → 799 and “nineteen ninety nine dollars” → $19.99 work.
• Indian amounts need their thousands word. “one lakh thirty five” stays as spoken; “one lakh thirty five thousand” → 1,35,000.
• Small counts convert only after a unit or count word. “for four people” → 4 people; “two nights” stays words.
• “point” is the noun after most words. “the margin is point five percent” → 0.5%, but “I take your point five minutes” → “I take your point 5 minutes”. Say “zero point five” after a verb or an article.
• “may” is the month only when the recogniser capitalised it. “on May second” → May 2nd; “on may second” stays, because “the second may be better” is the verb.
• Punctuation words two words after a determiner are commands. “a very long period comma” reads “a very long,”. Say “the word period”, or “full stop” for the command.
• Symbol words are written only in context. “hashtag micjot launch” → #micjot launch, but “hashtag activism is on the rise” keeps the word; “at the rate” is @ only inside an email address; “star sign” is never a symbol; a mid-line “dash” stays the word.
• Hyphen chains join every pair. “well hyphen known hyphen fact” → well-known-fact; say “well hyphen known fact”.
• Recogniser punctuation. “5 p.m. We leave” → “5 PM. We leave” when the recogniser capitalised the next word; a name, time zone or weekday after it gains no period.
• Other languages. With the language set to anything but English (or detected as such), spoken “comma” and “new line” are not turned into punctuation or line breaks; spacing and sentence capitals are still fixed.

With the AI step on, AI cleanup resolves most of these from context; see AI writing.

Accuracy and latency, measured

Recognition runs on your Mac. Synthetic-voice tests show a 0.54% word error rate on the benchmark's dev split (50 clips from sub-second words to 55-second passages, measured 2026-10-27); real voices, accents and headsets will score lower. A short dictation appears about one to two seconds after you release the key (median about 1.0 s, 95th percentile about 1.9 s). The first dictation after an install or update waits about a minute while Core ML compiles the model.