Dragon on the desktop is not Dragon Anywhere
Dragon has been the recognisable name in dictation for decades, and the desktop products for Windows and macOS do the recognition on the machine in front of you. Most people who know the name remember it from there: software that installed a large recognition engine locally, learned one speaker over time, and ran without needing anybody else's servers.
Dragon Anywhere is the mobile product. It is available on Android, and it is built differently. It works over a connection, and it wants an account and an active subscription before it will write anything down.
Neither of those is unreasonable for a product of that kind. It is simply not the thing people thought they were getting, and that mismatch is what sends most readers looking for something else.
LocalType is an Android keyboard where the speech model sits on the handset and does the recognition there, with no account and nothing transmitted.
Nothing here is affiliated with Nuance or Microsoft; LocalType is a separate product.
When a desktop licence lapses
The common version of this search is not somebody shopping around. It is somebody whose licence has run out, or whose work laptop went back to the company, and who is now wondering whether the phone in their pocket can take over.
That transfer does not really work, and it is better to know before you try. The desktop workflow rests on spoken commands, spoken formatting and a document window you can move around by voice. None of that exists in phone dictation of the modern kind, from any vendor, this one included.
What does carry over is the habit underneath it. Composing aloud in finished sentences, deciding what a paragraph says before opening your mouth, letting the text arrive and repairing it afterwards instead of picking at it mid-flow. That habit took months to build and it is worth more on a phone than any feature is.
So the question is not what replaces Dragon. It is which parts of your week were genuinely document work, and which parts were messages, email and notes that happened to be sitting on the same machine as the document work.
Commands, profiles and specialist vocabularies
The professional products earned their reputation, and the reasons they earned it have not gone away.
Spoken commands. Formatting, navigation, correction and selection by voice, which is what makes hours of hands-free document work possible. Modern phone dictation has none of that.
Voice profiles. Personalisation to one speaker, running over months, which pays off under heavy daily use.
Specialist editions. Legal and medical vocabularies carrying terminology a general model has never been shown.
Documents rather than fields. A working environment for producing long structured text, with the document living inside the app you dictate into.
If your work is two hours a day of reports with formatting spoken aloud, buy a professional dictation product and keep it on a computer. A phone keyboard is not a substitute for that, and nobody here is pretending otherwise.
A key on the keyboard, not an app to open
Dragon Anywhere is a place you go to dictate. LocalType is not a destination at all. The microphone is a key on the keyboard you are already typing with, so the words land in the field your cursor was in, with nothing copied and the clipboard left alone.
That covers messaging, email, notes, the browser, forms, search boxes, and the hundred short pieces of writing that make up a day on a phone. It works wherever a keyboard opens, on phones and tablets running Android 8.0 and newer.
There is also nothing to sign into. LocalType has no registration step and no profile, so the keyboard is not tied to an identity and there is no session that can expire on a platform at seven in the morning.
Sizing the engine yourself
In one respect a phone keyboard behaves more like desktop dictation software than like most phone software: you choose how big the recogniser is.
Three models ship with the app. Compact occupies 60 MB and returns text quickest, which is what an older or memory-starved handset actually needs. Standard occupies 190 MB and is what the app suggests unless you overrule it. Large occupies 539 MB and handles accents, quick talkers and background noise best, in exchange for patience. On a test phone with 4 GB of memory, Standard turned 6.9 seconds of speech into text in 4.1 seconds, while Large takes roughly twice as long as you spent speaking.
Only one of the three is kept on the device at a time. Downloading a different size replaces the file rather than adding to it, so trying all three costs you the download and nothing else.
Internet access is requested for exactly one purpose, collecting the model you picked, and the app does not reach outward again after that. A tunnel, a flight, a rural stretch of motorway and a hotel with a captive portal are therefore not situations the app has to cope with, because it was never depending on the signal in the first place.
Eighteen languages, none of them declared in advance
Recognition covers eighteen languages, and they all come out of a single multilingual model rather than a pack per language. Which of them you are speaking is worked out from the recording, so nothing has to be set before you start and nothing has to be switched between one message and the next.
Those eighteen are not equally strong. The material behind them is not equally plentiful, and the widely spoken languages benefit from that. Every speech model has the same shape of problem, and knowing about it beforehand is better than discovering it in the middle of a message.
Keyboard duty, and the words you add to it
On the days you never once tap the microphone it still has to be a keyboard you would willingly type on. It is: typos corrected as you go, the next word offered, a word list matched to the language in use, QWERTY and QWERTZ and AZERTY where those are the national standard, and a full emoji set with your recent picks at the front.
Names, jargon, brand names and internal acronyms can be added by hand. Those additions do double duty, because one app owns both halves of the job: the list feeds the typing suggestions, and it is handed to the speech model as context while you dictate. It is the closest thing here to the voice profiles desktop dictation is known for, though it works from the other end. Not adaptation to your voice, but adaptation to your vocabulary.
The rest of the housekeeping follows from keeping the work on the handset. The model file sits in storage no other app can open. There is no advertising anywhere in the product and no behavioural tracking inside it. Nothing keeps an archive of your recordings, because the sound becomes text and stops existing. The app's data is held out of Android's cloud backup on purpose, which is why a replacement handset starts blank and you fetch a model again. And when the cursor moves into a password field, the microphone key is not offered.
Two hours a day of structured documents is a professional dictation product and a computer. Forty short bursts a day, in whichever app you happen to have open, is a keyboard. Plenty of people run both, on different machines, and there is nothing contradictory about that.