A notepad against a keyboard
Speechnotes is a dictation notepad. You open it, you talk, the words appear in a document, and when you are done you copy them into wherever they were actually meant to go. Speechnotes has been around for years, it is free to use with advertising, and dictating in it does not require you to register at all.
LocalType is not a notepad. It is the Android keyboard itself, with a microphone key, so the text arrives directly in the field you were typing in. The speech model that recognises you runs on the phone.
The two have no connection to each other. LocalType is a separate product and is not affiliated with Speechnotes.
Two things Speechnotes does better
For a good number of people these two settle the matter.
It runs anywhere with a browser. A borrowed laptop, a desktop at work, a tablet you do not own, a machine in a library: open the page and dictate. LocalType has to be installed on a device you control, and it is Android only.
It is free to start with. LocalType is a paid product after a trial, and and there is no way to spin that as an advantage.
If your dictation happens at a desk, in long stretches, into a document you will edit afterwards, a notepad in a browser is the right shape for the job, and nobody should talk you out of it.
Whose recognition is actually listening?
Speechnotes is upfront about how it works. For dictation it uses the speech recognition built into your browser or your operating system, never one of its own, and for transcribing audio files it uses cloud engines from Google and Microsoft. That is a sensible way to build a product, and it is also why the question "where does my voice go" does not get answered by Speechnotes. It gets passed along to whichever engine your device happens to provide.
So if you came to a dictation tool hoping to move away from cloud speech recognition, a tool built on top of the platform's own engine has not moved you anywhere. It is a better notepad, not a different answer.
Our side of that question is answered directly rather than delegated. The recognition is a model file sitting in the app's private storage on your handset, no other app can open it, and there is no second engine involved and no service behind it to fall back to.
Copy, paste, fix the spacing
The workflow difference is easy to underrate until you have lived with it for a week.
With a dictation notepad the loop runs: leave what you were doing, open the notepad, dictate, read it back, select all, copy, return to the message, paste, fix the spacing where the paste landed badly. Every step is fine on its own. Together they mean dictation becomes something you do when it is worth the detour, and most of the time you thumb-type instead.
There is no detour here, because the microphone is a key on the keyboard that opened when you tapped the message box. You type, you press it, you speak, you carry on typing. Nothing is copied, nothing is pasted, and the clipboard still holds whatever it held before. A one-line reply gets dictated as readily as a paragraph, which is the case a notepad can never really serve.
Dictation works anywhere a keyboard opens, on phones and tablets running Android 8.0 or newer.
Where the first copy stays
Dictating into a notepad leaves your words in two places: the document you dictated into, and the message you eventually pasted them into. The first one usually stays there. Weeks later a browser tab or an app still holds a draft resignation letter, a doctor's appointment, a password hint you spoke aloud because typing it was awkward.
None of that is a flaw in Speechnotes. It is what a notepad is for, and the copy is the point of the product.
Our app simply does not work that way. The text lands in the field you were typing in and nowhere else, exactly as if your thumbs had produced it, and no archive of the recordings that produced it is kept anywhere in the app. The microphone also switches itself off in password fields, so no microphone key appears there and the most awkward category cannot be dictated by accident. There is no drafts folder to go back and clean out, because one was never created.
Nothing accumulates on the other side either. Speechnotes funds its free tier with advertising, and asks for an account once you move from dictating into transcribing files. There is no advertising in LocalType at all, so no advertising profile is being assembled in the background. There is no registration step, no login and no profile, so nothing ties the keyboard to a person in the first place. No behavioural tracking runs inside the product, and its data is deliberately kept out of Android's cloud backup, so a new phone starts completely fresh and inherits nothing at all.
Three sizes of the file that does the listening
Because the recognition is ours rather than borrowed from the platform, the model has to live somewhere, and that somewhere is your storage.
You get three sizes, so the decision is not made on your behalf. Compact takes 60 MB and returns text fastest, which is what matters on an older or memory-tight handset. Standard takes 190 MB and is the one the app suggests for most phones. Large takes 539 MB and reads accents and noisy rooms best, though it makes you wait: measured on a 4 GB phone, Standard turned 6.9 seconds of speech into text in 4.1 seconds, and Large takes roughly twice as long as you spent speaking. Only one is kept at a time, so trying a different size replaces the file instead of adding to it.
Fetching that file is also the one thing the app asks for internet access for, and it asks for exactly one purpose. A browser-based notepad needs a browser, a page load and, for anything beyond the platform engine, a working connection. After the download, LocalType needs none of it, so dictation carries on in a tunnel, on a flight, in a basement, or abroad with mobile data switched off.
Underneath the microphone it is a normal keyboard, and that matters on the days you never dictate at all: autocorrect and next-word suggestions from a dictionary matched to your language, the layouts you would expect including QWERTZ and AZERTY, a full emoji set, and your own names and jargon added by hand. Those added words improve typing and are given to the speech model as context, so they come out right when you say them too.
Eighteen languages, chosen for you per recording
LocalType covers eighteen languages with one multilingual model. It works out which of them you are speaking from the recording itself, so switching language between two messages needs no trip to a settings screen. Pin it to one language instead and the layout, the dictionary and the dictation language line up together.
Quality is not equal across all eighteen. Speech models are trained on very different amounts of material per language, and the widely spoken ones come out ahead.
Documents, or everything else
Speechnotes is for dictating documents: the long ones, at a desk, where you were going to edit afterwards anyway and the copy step happens once.
LocalType is for dictating everything else. The reply you are halfway through. The search box. The note to yourself while the kettle boils. The form field on a site that will not remember your address. The message you started typing and then thought better of thumbing out in full.