Why day one is the worst day
Three things are working against you at once, and none of them are the software.
Speaking to a phone feels absurd. Not technically, socially. You are performing an unfamiliar act, usually somewhere you can be overheard, and self-consciousness produces exactly the delivery that recognisers handle worst: quiet, hesitant, trailing off.
You have not learned the delivery yet. Complete sentences with pauses at the end is a skill, and on day one you are thinking aloud instead, which is the worst possible input.
Nothing knows your vocabulary. Every name you use is wrong, every time, because you have not added any yet. That single factor produces most of the errors people cite when they give up.
All three improve within days. The software does not change at all.
What to do before you start
Fifteen minutes here saves the entire experiment.
Download the model on Wi-Fi and confirm it works. A half-finished download that fails later gets blamed on dictation, and never on the download itself.
Add thirty words. Family, colleagues, your street, your town, the products and jargon you write about constantly. This is the single highest-return action available and almost nobody does it first. Anything you add stays on the phone and is passed to the recogniser as context.
Pick the middle model unless your phone is short of memory, in which case pick the smallest. Do not start with the largest and conclude dictation is slow.
Choose one thing to use it for. Replying to family messages, or notes to yourself, or nothing but email. Doing the same thing every day builds a habit; a general intention to try something does not.
Days one and two
Expect it to feel strange and produce mediocre results. That is the assignment.
Use it only for the one thing you chose. Do not attempt a long email, do not use it in front of colleagues, and do not judge anything yet.
The single adjustment to make immediately: hold the phone closer. Most day one complaints come down to distance, not to software, and it costs nothing to fix.
If a word is wrong, fix it and carry on. Do not stop mid-dictation to correct things; that breaks the rhythm and makes the rest worse.
Days three and four
The delivery starts settling here, and most of the improvement happens in these two days.
You will notice yourself finishing sentences rather than trailing off, and pausing at full stops without deciding to. Punctuation improves immediately as a result, because that pause is the strongest signal the system has.
This is also when to add the second batch of vocabulary: everything you have corrected twice in the first few days. That list is much better targeted than anything you could have guessed in advance.
If results are still poor, check one thing before anything else: whether the keyboard language matches the language you are speaking. A mismatch produces confident nonsense and no error message.
Days five to seven
By now the question changes from "does this work" to "where does this belong".
Most people find the same boundary. Anything longer than a sentence gets dictated. Short replies, addresses, numbers and codes get typed. That split is not a compromise; it is what the tool is actually good at.
This is the week to try it somewhere new. If you only used it for messages, try a note after a phone call, or an email you have been postponing. The value shows up most in the writing you were not doing at all.
When to conclude it is not for you
Dictation genuinely does not suit everyone, and four signs tell you early.
If you have nowhere to speak. An open-plan office, a shared home, a job in quiet spaces. This is the most common real reason and no amount of practice addresses it.
If your writing is mostly short and technical. Codes, numbers, references, single-word fields. Typing wins there and always will.
If speaking is effortful. Anything that makes talking tiring makes this tiring.
If after two weeks you are still correcting more than you are saving. Try a larger model first, and check the vocabulary. If both are done and it still does not pay, it does not pay.
What is not a good reason to stop: that it felt strange, that day one was bad, or that you noticed the errors more than the time saved. All three describe the first few days. None of them describes the tool.
The mistake almost everybody makes in week one
People try dictation first on the hardest possible message: the long, careful, important one they have been putting off. It is the message on their mind, so it is the one they reach for.
This is precisely backwards. A difficult message combines everything dictation handles worst on day one, unfamiliar delivery, high stakes, probably a name or two, and usually somewhere you feel self-conscious. It goes badly, and the conclusion is that dictation does not work.
Start with the messages that do not matter. Confirmations, replies to family, notes to yourself. Learn the delivery on writing where an error costs nothing, and arrive at the difficult message in week two with the habit already in place.
What most people report after a fortnight
Two changes, and neither is speed.
The messages they were postponing get written. Not faster than before, but at all. Shaving seconds off the things you already do is a smaller and much less interesting benefit than that.
And the split becomes automatic. They stop deciding whether to dictate; the hand goes to the microphone key for a paragraph and to the letters for a postcode, without any thought at all.
Somewhere around the fortnight mark, people stop noticing they use it at all. An input method you stop noticing is exactly the goal.
What is unchanged from day one
None of the settling-in affects the rest of the product. Recognition happens on the handset from a model in storage no other app can open, and internet access is used for exactly one thing, collecting the model you chose.
Nothing in it asks you to register. There is no advertising, no behavioural tracking sitting inside the product, and no archive kept of what you dictated. The app's data is held back from Android's cloud backup by design, and the microphone key stays away from password fields.
The upshot: the fortnight above can happen on a train, on a flight or in a basement, and nothing about it depends on you having signal while you learn.
If you stopped before
A good share of people reading this tried dictation years ago and put it down. The technology moved since then. Recognition on ordinary speech got substantially better, punctuation stopped needing to be spoken aloud, and models became small enough to run on a handset without a connection. Someone whose opinion was formed on the previous generation is describing software that no longer exists.
The habits above still apply, and the settling-in period is the same. What is different is that the ceiling is higher than the one you hit last time.