First check that typing is the problem
Half an hour spent installing the wrong answer is how a good idea gets a bad name, so start by ruling out the four situations where dictation is not the fix.
If reading the screen is the difficulty and typing is not. Dictation produces text that still has to be read. Android's own display settings, larger text and higher contrast, address that directly and cost nothing.
If hearing is the difficulty in conversations. That is a captioning problem, not a dictation one, and there is software built specifically for it.
If speech itself is effortful. Anything that makes speaking tiring makes dictation tiring, and for some people typing genuinely remains the easier route.
In a room with other people. Dictation is spoken aloud, so it does not suit a shared lounge, a waiting room or a care setting with company. It suits a kitchen table, a garden, a phone call's worth of privacy.
Ask about those four before you touch the phone. Discovering them afterwards costs somebody an afternoon and leaves them convinced that new technology is not for them.
Small keys and nothing to tell you that you missed
Where typing is the barrier, it has nothing to do with capability. Phone keyboards are simply badly suited to a lot of hands and eyes.
The keys are small and close together, which punishes any loss of fine motor control, and there is no tactile feedback to tell you whether you hit the right one. Rendered text is small too, so checking what you wrote means finding reading glasses. Autocorrect changes words confidently and wrongly, and undoing it requires knowing a gesture nobody explained.
What comes out of that is short messages, long delays, and eventually a phone used for calls only. None of that is a preference. It is friction.
Speaking a sentence takes a few seconds and no precision at all. There is no target to hit, no autocorrect fighting you mid-word, and no need to read anything while producing it. For somebody who has been sending three-word replies because each character costs effort, that is not a productivity improvement. It is the difference between participating in a conversation and dropping out of it.
Expect longer messages, sent sooner, with fewer abandoned halfway.
Do the setup on their phone, not yours
Handing over an installed app and explaining it is the mistake that decides the whole outcome.
Do the whole setup on their phone yourself. Enable the keyboard in Android's settings, select it, download the speech model on Wi-Fi and confirm dictation works before you hand the phone back. Anything left half-configured will simply not be used.
Add the names first. Family, grandchildren, the street, the town, the doctor, the neighbours. These are exactly the words a general speech model has never encountered, and they are also the words that will appear in every message. LocalType keeps those additions on the phone and gives them to the speech model as context, so they start arriving spelled correctly.
Choose the model deliberately. A quiet room and clear speech does not need the largest one, and on an older handset the smallest will feel quicker. LocalType offers three, at 60, 190 and 539 MB. The middle setting turned 6.9 seconds of speech into text in 4.1 seconds on the 4 GB test phone.
Leave the old keyboard enabled. Switching between keyboards is one tap, so nothing is lost if the new one turns out not to suit them. Show them where that switch is.
Then demonstrate once, and let them try immediately. Watching somebody do it teaches nothing. Doing it once with you sitting there teaches everything.
Four things to explain, and the list not to
Where the microphone key is. That you press it, speak a whole sentence, and stop. That the words appear where you were typing. And that if a word comes out wrong, you fix it exactly the way you would fix a typo.
Do not explain models, offline behaviour, privacy architecture or settings. None of that is needed to use it, and a ten-minute explanation is how people conclude the thing is complicated.
Two warnings belong with those four, because unexplained surprises are what kill new habits.
It will get names wrong sometimes. Especially unusual ones, especially early on. Nothing is broken when that happens, it is not something they did wrong, and it improves as you add words.
It writes exactly what was said. Including "um", including a restarted sentence. It is not tidying anything up, so speaking in complete sentences and pausing at the end produces much better results than thinking aloud does.
That second point lands far better demonstrated once than described twice. Say a rambling sentence into the phone, let them read what appears, then say a clean one and let them read that.
Thick walls and small data bundles
The offline part matters in this particular case, and not for reasons of principle.
Older people are disproportionately on cheaper contracts with smaller data allowances, and dictation that uploads audio consumes them. On-device recognition sends nothing at all.
And they are disproportionately in the places where signal is poor: hospitals, care settings, older buildings with thick walls, rural areas, the middle of a supermarket. Dictation that stalls in those places gets blamed on the phone and then abandoned.
LocalType runs the model on the handset, so behaviour does not change with coverage. Internet access is requested for exactly one purpose, collecting the model you chose, and never again. That also means there is nothing to sign into, which removes an entire category of the calls that begin "it is asking me to log in again".
The privacy side needs one sentence for the family member and none at all for the person using it. Nothing is transmitted and no account exists. There is no advertising, no behavioural tracking runs inside the product, and no archive of the recordings is kept. Only the app can reach the storage the model sits in. Android's cloud backup is deliberately denied the app's data, and in a password field the microphone key does not appear. That last one is practical rather than philosophical: it stops a password being dictated by accident into a login screen.
Two weeks, one job
Speaking to a phone feels strange, and people who have typed for twenty years will revert to typing several times before it settles.
A specific prompt makes the transition happen faster. Suggest they use it for one particular thing, replying to family messages, and only that, for two weeks. Habits form around one small repeated job far more reliably than around a general instruction to try something.
Check back after that fortnight and add whatever names have been coming out wrong. That second ten-minute session is usually what turns it from an experiment into the way they write.
Four lines on a piece of paper
Write down those four things on a piece of paper and leave it by the phone. Not a manual, four lines.
People who would never open a help page will happily glance at a note, and it removes the phone call that starts with "I have forgotten how to do the thing you showed me".