← Back to Blog

Short answer: Dictation converts numbers to digits based on context, not on how you say them. Speak numbers inside a full sentence, say a number's natural spoken form rather than digit-by-digit, and say symbol names like "percent" and "dollars" as words. Then proofread every figure, because number errors are silent.

Most dictation frustration is about words. This one is about figures, and it behaves by completely different rules. You say "twenty five" expecting 25 and get "twenty-five." You say a phone number and get a paragraph of spelled-out digits. You dictate an invoice line and the currency symbol lands on the wrong side of the amount.

None of that is random. Speech recognition treats numbers as a formatting decision made after the words are recognized, and that decision is driven almost entirely by the surrounding context. Once you know what the engine is looking at, you can feed it what it needs and get clean figures on the first pass. Here is the full set of rules, surface by surface, plus an honest list of the numbers you should not be dictating at all.

Why numbers come out spelled instead of as digits

When you speak, the recognizer produces a sequence of words first. "Twenty five" arrives as two ordinary words. A second stage then decides how to render them in text, and that stage follows conventions borrowed from written English rather than from what you literally said.

Written English has a long-standing convention: small numbers get spelled out, large ones get digits. Style guides differ on where the line sits, but somewhere around nine or ten is common. Dictation engines learned from enormous amounts of written text, so they inherited that habit. When you say "I have three meetings," you get "three," not "3," because that is how the sentence would have been typed by a person.

This is the single most useful thing to understand: the engine is not converting your speech to digits, it is writing the sentence the way a competent writer would. Fight that and you will lose. Work with it and you get clean output.

The corollary is that context flips the decision. The same spoken word renders differently depending on what surrounds it:

Notice that in every case where digits appeared, there was a signal nearby: a label, a unit, a symbol word, a month. That is the lever you pull.

The core rule: give the number a job

A bare number floating in a sentence is ambiguous, so the engine defaults to prose style. A number attached to a unit, a label, or a symbol is unambiguous, so the engine formats it as data.

Practically, this means you should stop trying to force digits and instead say the thing the digits describe. Do not say "five" and hope. Say "five percent," "five dollars," "five kilometers," "version five," "page five," "five p.m." Every one of those gives the formatter something to anchor to.

When a number genuinely has no unit and you still want digits, the reliable approach is to dictate the sentence and fix that one token afterward, or to set up a replacement rule so it happens automatically. More on that below.

Large numbers: say them the way you would read them aloud

The instinct when a number gets long is to slow down and read it digit by digit. This is exactly backwards. Digit-by-digit speech strips away the structure the engine uses to assemble the number, and you often end up with a string of separate small numbers instead of one large one.

Say the number the way you would read it to a colleague:

Grouping separators generally follow your locale's convention automatically. If your output uses the wrong separator style entirely (periods where you expect commas, or vice versa), that is a language-variant mismatch rather than a number problem, and it is worth checking your dictation language setting.

One reliable trap: "one hundred and five" versus "one oh five." The first gives you 105. The second is ambiguous and often renders as separate tokens, because "oh five" reads as a sequence rather than a quantity. Use the full spoken form for quantities and reserve the "oh" pronunciation for identifiers where a sequence is what you actually want.

Decimals and fractions

Say "point" for a decimal separator. "Three point one four" gives you 3.14. Do not say "dot," which is a URL and filename word and may render as the literal character in the wrong place.

For currency amounts, the natural spoken form usually beats the technical one. "Nineteen ninety nine dollars" is more likely to be misread than "nineteen dollars and ninety nine cents" or "nineteen point nine nine dollars." The spelled-out cents version is verbose but it is unambiguous, and unambiguous is what you want on anything with money in it.

Fractions work best as their common spoken names: "one half," "three quarters," "two thirds." These usually render as words rather than as fraction characters, which is normally what you want in prose. If you need a typeset fraction or a specific glyph, type it, since that is a formatting task, not a dictation task.

Percentages, degrees, and units

Say the unit as a word and let the engine decide on the symbol. "Twelve percent" typically renders as 12%. "Twenty degrees Celsius" gets you the degree notation. "Five kilograms" and "ten miles per hour" behave the same way.

Where this breaks down is with abbreviations that are also ordinary words. "Five in" for five inches will not do what you want, because "in" is a preposition. Say "five inches." Same for "two ft" versus "two feet." Always speak the full unit name; the abbreviation is the formatter's job, not yours.

Ranges are one of the more reliable constructs. "Between ten and fifteen percent" comes out cleanly. "Ten to fifteen percent" usually does too, though "to" occasionally renders as the word rather than a dash. If you specifically want an en dash range, say the numbers and insert the dash yourself.

Dates: the highest-accuracy number you will ever dictate

Dates are the good news in this article. Because month names are unmistakable anchors, date dictation is genuinely reliable. Say the date the way you would say it out loud and it lands:

Two things to watch. First, ordinals: say "third," not "three," when you mean a day. Saying "March three" often produces exactly that, and it reads wrong. Second, numeric date formats. If you need 03/15/2026 specifically, dictating it is a losing battle because slashes in spoken form are unreliable and the day/month order convention varies by locale. Say the date in words, or type the numeric form.

For scheduling text specifically, times behave well. "Two thirty p.m." gives you 2:30 PM in most contexts. "Half past two" stays as words. "Fourteen hundred hours" renders as spoken. If you are dictating meeting notes or calendar entries, the p.m./a.m. form is the one to standardize on.

Phone numbers, order IDs, and other sequences

This is the category where dictation genuinely struggles, and it is worth being honest about why. A phone number is not a quantity. It is a sequence of independent digits with no linguistic structure, which means the engine has nothing to work with. It cannot use context, because there is no context inside a random string.

You can improve your odds:

But the real answer is the same one that applies to dictating email addresses and URLs: stop dictating strings you type repeatedly. If you say your own office number, your support line, or your standard order-reference prefix more than twice a week, a replacement rule turns a spoken shorthand into the exact string, every time, with zero recognition risk. That is not a workaround; it is the correct tool for the job.

Numbers you should never dictate

Some figures should not go through a microphone at all, regardless of accuracy. The reason is not recognition quality — it is that you are saying them out loud in a room, and speech has an audience that typing does not.

Do not dictate:

Type those. The accuracy risk on a long unstructured string is high, the consequence of a silent digit error is severe, and speaking a credential aloud in an open office or a cafe is a disclosure whether or not anyone was listening on purpose. If you want the fuller picture on what happens to audio in a dictation pipeline, we covered it in does voice to text store your audio.

Why number errors are more dangerous than word errors

A misrecognized word usually announces itself. The sentence reads oddly, your eye snags, you fix it. A misrecognized number does nothing of the kind. 1,500 and 15,000 are equally plausible in an invoice. March 3 and March 13 are equally plausible in a meeting note. Nothing in the sentence looks wrong.

This is the reason the workflow advice for numbers differs from the advice for prose. With prose you can dictate a long stretch and clean up at the end. With figures you should verify at the point of entry, because the error will not surface later on its own.

A practical habit: dictate the sentences, then do a separate pass over the numbers only. Read each figure against its source. It takes twenty seconds on a page and it catches the entire class of silent errors. The same two-pass approach shows up in most of the professional workflows we have written about, from dictating in QuickBooks to project tracking, and for the same reason: prose and data want different kinds of attention.

The boundary that saves the most time

The single most useful rule across every app we have covered is this: speak the sentences, type the values.

Dictation is extraordinary at prose. You speak at 130 to 150 words per minute, an average adult types around 40, and even strong typists land in the 80 to 100 range. That gap is where all the benefit lives, and prose is where the gap applies.

Values are the opposite. A dollar amount is four keystrokes. A date picker is two clicks. A quantity field is one number. There is no speed gap to exploit, and there is a real accuracy cost. Reaching for the mic on a numeric field is spending risk to save nothing.

So the efficient workflow in almost any form-shaped app is: dictate the description, the note, the summary, the explanation, the email — then type the figures. That is not a limitation of voice. It is voice being used where it wins.

Where Voice Keyboard Pro fits

Voice Keyboard Pro is a voice-to-text app for Mac and iPhone. On Mac it lives in the menu bar: hold a hotkey, speak, release, and the text appears at your cursor in whatever app is in front of you. Because the text is inserted at the cursor rather than through an app-specific plugin, it works in a spreadsheet cell, a browser form, an invoicing tool, and a code editor without any per-app setup.

Two features matter specifically for the number problem.

Smart Vocabulary is a personal dictionary with replacement rules. This is the real fix for recurring strings. Instead of dictating your standard reference prefix, your office number, or a product SKU pattern digit by digit and hoping, you say a short phrase and the exact string is inserted. Recognition risk drops to zero on precisely the content where recognition is weakest. It also handles the proper-noun problem (client names, product names, internal system names) that no microphone upgrade can solve.

Voice Edit, on the iPhone keyboard, lets you speak a correction rather than fighting a text cursor on glass. When a figure comes out wrong in a message you are drafting on your phone, saying what should change is faster than tapping to position a caret between two digits.

The iPhone side is a custom keyboard with a built-in mic button, so it works in any iOS app, and it includes two-way translation across 24 languages while dictating along with swipe typing. There is a free tier with daily limits; Pro is $4.99 a month or $34.99 a year. Our server stores only operational pings: no audio and no transcript content.

Quick reference

Keep this near your desk for the first week:

If your figures are coming out wrong in a way none of this explains, the problem may be upstream of formatting. Recurring garbled output usually traces back to the audio path rather than the number logic, and the background noise fix guide covers the microphone and environment side in detail. And if your dictation is inserting punctuation you never asked for around your figures, the random commas guide explains why that happens and what to do about it.

Numbers will never be the part of dictation that feels magical. But they stop being the part that costs you time once you speak them in their natural form, let the engine handle the formatting, and reserve the keyboard for the handful of values where typing was always going to be faster anyway. Try Voice Keyboard Pro free and put the rules above through a real document.