Judged against what, exactly
People new to dictation measure it against an imaginary flawless typist, decide it falls short, and go back to the keyboard. That is the wrong comparison generally. For this group it is badly wrong.
The realistic alternative is not perfect typing. It is typing that hurts, or typing broken up by rests, or not sending the message at all, which is the outcome nobody counts because nothing appears on a screen to be counted.
Against that, a system that mishears a word now and then and needs a correction comes out ahead. Two fixes on a four-line message is a different afternoon from four lines picked out with a thumb that was already sore at nine in the morning.
Sore hands, tremor, spelling, low vision, fatigue
Sore or unreliable hands, whether from repetitive strain, arthritis, an injury or a condition affecting grip and fine movement. Phone keyboards ask for precision in a small area, which is exactly the demand that hurts.
Tremor, which makes small targets unreliable and turns a two-line message into an exercise in correction.
Spelling difficulty, including dyslexia, where saying a sentence is straightforward and writing it is not, and where autocorrect often guesses wrong and makes matters worse.
Low vision, where the difficulty sits in reading the keyboard rather than in pressing it.
Fatigue and pain conditions, where the problem is not any single message but the accumulated cost of a day of them.
None of that needs a diagnosis to be worth solving, and plenty of people simply find typing on glass unpleasant. Dictation is sold as a convenience. For everyone above it is the input method, and the marketing rarely speaks to them.
Four gains, and where each one stops
Effort per message drops, and corrections still need hands. Speaking a sentence takes a few seconds and no fine motor control, which for anyone rationing hand use across a day is not a comfort but a budget. You will still touch the screen to place the cursor and repair a word. There is less of it.
Length stops costing anything, and speech starts costing something. People who find typing difficult write shorter messages, not because they have less to say but because every character has a price. Dictation takes that ceiling away. In exchange it wants a reasonably quiet room and a setting where speaking aloud is appropriate, so it is no use on a crowded bus or in a shared waiting room.
Spelling stops being the gate, and names become the weak point. You say the word, the system writes it. For anyone who knows exactly what they mean and finds the written form the obstacle, that is the whole problem solved. What a general model gets wrong most often is proper nouns, which is the material where a mistake carries the most weight, so proofreading does not go away.
The screen matters less, and it does not vanish. Opening the app, placing the cursor, tapping the microphone, reading the result back. Screen contact goes down a long way, not to zero.
One limit sits outside that list because nothing offsets it. Dictation requires speech, which rules it out for some people entirely, and it works less well for anyone whose speech differs from what the model was trained on, including many speech differences that have nothing to do with accent. That is a real limitation of the technology, and you want to know it before you give a week to the method.
Five questions that matter more here than for casual use
Does it work everywhere? If typing is difficult, an input method that only works inside one app is not an input method. LocalType is a keyboard, so it appears wherever a keyboard appears, across the whole surface of the phone.
How many taps to start? For someone rationing hand use, the gap between one tap and six decides whether dictation gets used at all. Here it sits on one key of the keyboard already on screen, with no separate application to open first.
Can you add your own vocabulary? Names of family, carers, medication and places are exactly the words a general model gets wrong, and correcting them by hand every time is the cost you were trying to avoid. You can add your own words, they stay on the handset, and they are handed to the speech model as context, so the names you use most start arriving right.
Does it need a connection? Not a philosophical question in this context. If dictation is your input method, it failing in a lift, a hospital basement or on a train means the phone stops working for you. Recognition runs on the handset from a model file kept in storage no other app can open, so signal strength changes nothing. Internet access is requested for exactly one purpose, collecting the model size you picked, and afterwards the app has no reason to ask again.
What does it want from you in return? Dictation on a phone usually costs something that is not money. There is no account here, so nothing is attached to a name. No advertising sits in the product, nothing inside it keeps a record of how you use it, and none of your recordings is archived. The app's data is held out of Android's cloud backup on purpose. Where the cursor lands in a password field, no microphone key is offered.
Screen readers, untested and therefore unclaimed
LocalType has not been tested against screen readers or other assistive technologies, so this page makes no claim about how it behaves alongside them.
If you rely on those tools, try any keyboard against them yourself before you depend on it, ours included. An afternoon with your own setup tells you more than a product page can.
Choosing a model size, then choosing your twenty names
Three model sizes are offered, at 60, 190 and 539 MB. That is a trade between speed and accuracy rather than a ladder you should always climb to the top of. If dictation is doing most of your input, the largest is often worth the wait, because every avoided error is a correction your hands do not have to make. On the 4 GB test handset the middle size turned 6.9 seconds of speech into text in 4.1 seconds, and the largest takes roughly twice the time you spent speaking.
Eighteen languages run from one multilingual model, and the language is worked out from the recording rather than chosen from a menu, which spares anyone who writes in two of them a trip into settings.
For the first sitting, spend five minutes on the twenty names you use most: family, carers, the two or three medications, your street, the surgery. Then use it for a day on the messages you would otherwise have left until evening, and keep a note of any word you had to fix twice. Those go into the vocabulary next.
Give it a week rather than an afternoon. Much of what goes wrong on the first day is delivery rather than software: a sentence begun before the key was pressed, an ending swallowed, a run-on that should have been three sentences. Those are things you can change on purpose, once you know to watch for them.