The part everybody quotes
People speak several times faster than they type on a phone. That gap is real, it is large, and it is why dictation feels dramatic the first time it works.
Thumb-typing is slow because it is a precision task on a small surface with two digits. Speaking is something you have done all your life without thinking about it. Whatever the exact figures for you personally, the ratio is not close.
If producing words were the whole job, the question would be settled in one paragraph.
Time speaking plus time correcting
It is not the whole job, because output arrives with errors and errors cost time.
The real comparison is not speaking speed against typing speed. It is:
time speaking + time correcting against time typing + time correcting
Typing has errors too, and autocorrect handles many of them silently. Dictation has a different error profile: fewer typos, more wrong words, and the wrong words are real words. Nothing underlines them for you, so the noticing is your job.
That noticing is the hidden cost. A misheard word in the middle of a paragraph is invisible until you read the paragraph, and a paragraph you have just spoken is the hardest thing to read attentively, because your memory of what you meant is still fresh enough to paper over what is on the screen.
Run the test on your own writing
There is a version of this question that has an actual answer for you specifically, and it takes an afternoon in the background rather than a study.
Take a message you would genuinely send, of the length you usually write. Type it, timing from opening the app to being happy with it. Then dictate an equivalent one, timing the same way, including reading it back and fixing what is wrong.
Do that on three different kinds of message: a one-line reply, a paragraph, and something with several names in it. Three answers come back, and they will not agree with each other.
One caveat on the method. Do it after a week of using dictation, not on day one. Speaking in finished sentences rather than thinking aloud is a skill, it takes a week or two to pick up, and anybody timing themselves on the first afternoon is timing a beginner rather than a tool.
Long and familiar, or short and full of names
Two variables decide nearly every case: how long the thing is, and how many proper nouns are in it.
Length works for dictation. The speed advantage compounds as a message gets longer, while the correction cost tracks the number of unusual words, which does not grow at the same rate. Ordinary sentences made of ordinary vocabulary are the easy case and produce almost nothing to fix.
Proper nouns work against it. Names, addresses, part numbers, codes, anything dense with them. Exactly the material recognisers get wrong, and exactly where an error is expensive rather than embarrassing.
Very short things go to typing. A three-word reply, a search term, a postcode. The overhead of tapping the microphone, speaking and glancing at the result is fixed regardless of length, so on a three-word message it is the whole of the cost.
Unusual formatting goes to typing. Lists, tables, anything where the structure matters more than the words.
Two situations override the arithmetic entirely. When your hands are busy or unavailable, walking, cooking, carrying something, the comparison is not against typing but against not writing at all. And for anyone with limited hand use or pain when typing, the sum is not close and the correction cost is worth paying many times over.
Two more override it the other way. In company, where speaking aloud is not an option, and when you have not decided what to say yet. Typing gives you thinking time; dictation punishes hesitation with worse punctuation and half-sentences.
Somewhere around a sentence
Most people converge somewhere similar after a few weeks: dictate anything longer than about a sentence, type anything shorter.
That rule is not about the technology. Around a sentence is where the fixed overhead of starting a dictation stops mattering relative to what you save, and that boundary sits in roughly the same place regardless of which app you use.
Two levers inside the app
Two things sit inside a product's control and outside yours.
How fast the text arrives. For on-device recognition this depends on your hardware and the model size. Larger models are more accurate and slower, and a wait after every sentence is what makes dictation feel sluggish even when it is faster overall. LocalType offers three sizes, at 60, 190 and 539 MB, so you can position yourself on that line. On a phone with 4 GB of memory the middle size turned 6.9 seconds of speech into text in 4.1 seconds, and the largest takes roughly twice as long as you spent speaking.
How much overhead there is before you can speak. A keyboard with a microphone key has almost none: the keyboard is already open because you tapped a text field. A separate dictation app adds opening it, waiting for it, then selecting, copying, switching back and pasting, and on short messages that sequence is the entire difference between the two approaches.
There is also a category of delay that has nothing to do with either: waiting for a server. Cloud dictation adds network latency to every sentence and stops working altogether when the connection does. On-device recognition has no network in the loop after the initial model download, so the speed you get in a tunnel is the speed you get at home.
Four ways to shrink the correction half
Since corrections are what decides the answer, they are where to spend effort.
Add your own vocabulary. Names and jargon are the majority of dictation errors for most people. Fix them once and they stay fixed, which beats repairing them in every message. Those additions stay on the handset, and the speech model is handed them as context.
Speak in complete sentences. Recognisers use surrounding words to choose between similar-sounding options, so isolated words are the worst case you can hand them.
Hold the phone where it can hear you, and get out of the wind if you are outside. Noise mixes into the recording at the moment it is made and cannot be taken back out afterwards.
Match the model to the task. Short replies favour speed, long messages favour accuracy, and swapping between the two sizes replaces the file rather than adding to it.
Messages you would not have written at all
One effect never shows up in a stopwatch comparison.
The reply that needed three paragraphs and therefore waited until this evening. The note you would have trusted yourself to remember. The email you have been carrying around since Tuesday. The address you meant to send your sister before she asked twice.
Per message, dictation is somewhat faster. Across a week, what changes is which messages get written at all, and that is not something the three-message test above can measure. It is also why the very short replies, where typing genuinely wins, matter less than the arithmetic suggests: nobody was postponing those in the first place.