A browser tab against a keyboard
Dictation.io is a browser tool. You open it in Chrome on a Windows, Mac or Linux machine, you talk, and the words appear on the page, with spoken commands for punctuation and paragraphs. It handles well over a hundred languages, and what you dictated stays in your browser where you left it, not in an account somewhere.
LocalType is an Android keyboard. The microphone is a key on it, the words go straight into whatever app you were typing in, and the speech model doing the recognition is a file on your phone.
As a product, LocalType is separate and not affiliated with dictation.io.
"Stored locally" and "recognised locally" are different things
This distinction has to be right, because the two phrases sound alike and mean nothing like each other.
Dictation.io keeps your text in the browser, which is a genuinely good property: your notes are not sitting in some account you forgot you made. But the recognition is a separate matter. Google's speech recognition does the work, and the site lists an internet connection as a requirement, which tells you where the audio is being turned into words.
So the text is local. The listening is not.
LocalType inverts that. The model that converts your voice into text is downloaded to the phone once and runs there afterwards, and so the app works with the connection switched off entirely. Your audio does not need to travel to be understood.
If you have been choosing dictation tools on the basis of "it does not store my stuff", go back and check which half of the sentence you were being sold.
Three things the browser tool wins on
For some readers these end the comparison before it starts.
It needs no installation. Open a browser and start. On a machine you do not own, or one you are not allowed to install software on, that is decisive.
It is free and it covers more languages. LocalType is a paid product after a trial, and eighteen is a smaller number than a hundred. Neither of those facts improves by being argued with.
It takes spoken commands. You can say the name of a punctuation mark or ask for a new paragraph, and it obeys. That is a real feature for anyone dictating formatted text at length, and LocalType does not claim it: what our model produces is a transcription of the words you actually said, dropped into the field, and the punctuation it adds is what the model inferred, and you cannot dictate it by name.
Whether that matters depends on what you write. For messages and short replies, speaking punctuation aloud is more effort than typing a comma. For a long structured document, it is the difference between dictating and editing afterwards.
So if your dictation happens at a desk, in Chrome, in long stretches into a document you will edit afterwards, a browser tool is the right shape and there is no reason to move. The other case is dictating on a phone, in the app you are already in, without your voice going anywhere.
Copy, paste, and the laptop you did not open
Two problems come with a browser tab, and they compound each other.
The first is the device. A Chrome page lives on a computer, and most dictation does not happen at a computer any more. You are replying on the train. You are adding to a list in the kitchen. You are answering a message while walking. None of those is a moment where you open a laptop, load a website, dictate into it, select the result and carry it across to where it was meant to go.
The second is that last step, which never goes away. Select, copy, switch app, paste, tidy the spacing. Small, and it is the reason most people give up and type instead.
A keyboard is available in all of those moments by definition, because it is what appears the instant you tap a text field. That single fact is the reason LocalType is built as a keyboard and not as an app you visit. It works wherever a keyboard opens, in messaging, mail, notes, browsers and forms, on phones and tablets running Android 8.0 or newer.
There is no last step either. The words arrive in the field you were already in, exactly as if you had typed them, and the clipboard is never touched, so whatever you copied earlier is still sitting there when you want it. Nothing is left behind: no page holds a draft you dictated last Tuesday, and the app keeps no archive of the recordings.
In password fields the microphone deliberately refuses to work, which removes the one category where an accidental dictation is genuinely painful.
The weight of carrying your own model
Running recognition yourself means carrying the model, and that cost is real.
Three sizes ship with the app so the decision belongs to you rather than to us. The 60 MB one is the sprinter, and on a phone with little memory to spare it is the only sensible choice. The 190 MB one is the recommendation for most handsets. The 539 MB one hears accents and crowded rooms best but makes you wait for it: on a test phone with 4 GB of memory, the middle model turned 6.9 seconds of speech into text in 4.1 seconds, while the largest takes about double the time you spent talking.
Only one lives on the device at a time, so changing size swaps the file instead of stacking a second one beside it.
A browser page has none of that weight, which is genuinely an advantage until the moment you have no signal. Then the weightless option stops working and the heavy one carries on.
What you get back is that connectivity stops being part of the equation. Aeroplanes, tunnels, basements, foreign trips with data turned off: none of them are special cases, because nothing was ever being sent.
Eighteen languages, without a menu
Eighteen languages come from a single multilingual model in LocalType, and it works out which one you are speaking per recording, so you never set it first. Move from one language to another between messages and it comes along. Set it explicitly instead and the keyboard layout, the dictionary and the dictation language all follow that choice together.
Accuracy is not the same in all eighteen. Models see far more material in some languages than others, and the widely spoken ones benefit from it. Anyone promising uniform accuracy across a hundred languages is describing an ambition. No measurement backs it.
The trade for putting an app on your phone
A web page is free to open, and free things on the web are usually funded the same way. LocalType is not, and what that buys is the actual argument for putting an app on your phone instead of a tab in your browser.
There is no advertising in it, which also means there is no advertising profile being assembled from what you dictate. Nothing inside the product tracks what you do with it, and there is no account to create, so nothing ties the keyboard to a name. The model file lives in storage only LocalType can read, and the app's data is deliberately kept out of Android's cloud backup, so it is not copied off the handset that way either. Internet access is requested for one purpose, fetching the model you chose, and never for anything else.
Words you add yourself, colleagues' surnames, product names, the acronym your team uses daily, stay on the device and are handed to the speech model as context, so they start coming out spelled correctly.
And underneath all of it there is still a keyboard: autocorrect, next-word suggestions, a dictionary per language, the standard layouts including QWERTZ and AZERTY, and a full emoji set. You are not installing a dictation tool. You are installing the thing you type with, which happens to listen when you ask it to.