Five phrases that sound like guarantees
These are the ones that cost people money.
"Private." Undefined, and defined by nobody. Ask where the audio goes and whether an account is involved.
"Secure." Usually describes how a transmission is protected. Whether any transmission happens at all is a separate question and the more interesting one.
"Works offline." Sometimes means your text is stored locally, which says nothing whatever about where your voice was processed. Two entirely different claims wearing the same two words.
"Powered by Whisper." Says which model, not where it runs. An open model on somebody else's server is still on somebody else's server.
"No data collection." Narrower than it sounds. Ask what counts as data in that sentence, and whether the audio itself is included in the count.
Behind all five sits one test that does not care what the store listing says: turn off mobile data and Wi-Fi, then try to use the feature. The rest is the vocabulary you need to read the answer.
Recognition, dictation, transcription, captioning
Speech recognition, or ASR. Software that turns recorded sound into written words. ASR stands for automatic speech recognition and means the same thing. It does not understand what you said; it works out which words were spoken.
Dictation. Using speech recognition deliberately, to produce text you intended to write. Distinct from transcription in that you are the speaker and you are doing it on purpose.
Transcription. Turning an existing recording into text. Usually somebody else's speech, usually after the fact, usually producing a document at the end.
Captioning. Turning speech into text in real time so it can be read as it happens, typically for accessibility.
Marketing uses those four interchangeably, which is why a search for any one of them returns products built for the other three.
On-device, cloud, hybrid, and the mode that implies another mode
On-device, also called local or edge. The model runs on your phone or computer. Nothing is transmitted, and it works with no connection.
Cloud, also called server-side. Audio is sent to a remote server, processed there, and text comes back. Allows much larger models and requires a connection.
Hybrid. Both, with something deciding which one applies at any given moment. That decision may be yours or the software's, and it may depend on your device, your language or a setting you last looked at a year ago.
Offline mode. A hybrid product's local option. If there is a mode for it, there is by definition another mode sitting next to it, and something somewhere chooses between them.
Weights, quantisation, and why a phone can run this at all
Model. The file that does the recognising. Produced by training and essentially a very large collection of numbers.
Weights. Those numbers. When people say a model has "open weights", they mean the file can be downloaded and run by anybody, which is what makes on-device products possible in the first place.
Training data. The recorded speech and matching text a model learned from. The amount that existed differs a lot from language to language, and results follow it.
Model size. Usually quoted in megabytes. Larger models handle accents, noise and unusual words better, and need more memory and more time.
Quantisation. Storing a model's numbers with less precision so the file is smaller and faster to run, at a modest cost in accuracy. It is much of the reason phone-sized models became practical rather than theoretical.
Whisper. A speech recognition model released by OpenAI with its weights available. A great many on-device products are built on it, or on reimplementations of it.
whisper.cpp. A version of Whisper written to run efficiently on ordinary processors, widely used in mobile apps because most phones have no graphics card worth using.
Latency, accuracy, and the numbers people quote
Latency. The delay between you stopping and the text appearing. On-device latency depends on your hardware and model size; cloud latency depends on your connection.
Accuracy, or word error rate. How often it gets words wrong. Quoted figures almost always describe clean recordings in well-resourced languages, so treat them as a ceiling. Your own results will sit somewhere below it.
Real-time factor. How long processing takes relative to how long you spoke. Below one means faster than speech; above one means you wait. A figure like "6.9 seconds of speech in 4.1 seconds" is this, given as two times instead of a ratio.
Megabytes for a model. Storage occupied on the device. Memory used while running is a separate figure and usually a larger one, so a model can fit on a phone comfortably and still be uncomfortable to run on it.
Languages supported. A count of the languages a product will attempt, which is not the same as the languages it handles well. Comparing a hundred against eighteen tells you almost nothing about the one language you write in.
Far-field and near-field. Whether the microphone is across a room or close to your mouth. Dictation products are built for near-field. Far-field is a different design problem, aimed at rooms rather than at one person writing, and the two are not interchangeable.
Voice activity detection. Working out when speech starts and stops, so the system knows what to process and when you have finished talking.
Language detection. Working out which language is being spoken, with nobody telling the system in advance. Reliable on full sentences, unreliable on single words.
Inferred punctuation. Commas and full stops decided by the system from your rhythm and your grammar. Nobody has to say "comma" out loud. Standard in modern products, and the reason dictated text no longer arrives as one long clause.
Input method, wake word, personal dictionary
Input method, or IME. The technical name for a keyboard on Android. An input method receives your keystrokes and can read the text around the cursor, which is how autocorrect works and why Android warns you when you enable one.
Enabling versus selecting. Two separate steps. Enabling permits a keyboard to exist as an option; selecting makes it the one in use. Missing the second is the most common reason a newly installed keyboard appears to do nothing at all.
Personal dictionary. Words you add yourself so they stop being corrected and, in products that support it, so the recogniser knows they are possible words to hear.
Wake word. A short phrase a device listens for continuously in order to start an assistant. Dictation does not use one, and a keyboard does not listen between uses.
Cloud backup. Android's routine copying of app data to your account. An app can opt out, which protects the data and means it does not come back on a new phone.
LocalType in these terms
LocalType is an input method with on-device ASR, no hybrid mode, no wake word and no account. The model has open weights, is quantised to run on a phone, and is offered in three sizes at 60, 190 and 539 MB, so you choose your own point on the latency and accuracy curve. On the 4 GB test handset the middle setting turned 6.9 seconds of speech into text in 4.1 seconds, and only one size sits on the phone at a time.
Language detection covers all eighteen languages, taken from the recording rather than from a setting, and punctuation is inferred from your rhythm instead of dictated aloud as commands. A personal dictionary is supported, stays on the handset, and is given to the model as context.
Internet access is requested for exactly one purpose, collecting the model. Advertising is absent, no behavioural tracking runs inside the product, no archive of your recordings is kept, the model sits in storage only the app can reach, and the app's data is kept out of Android's cloud backup. Tap into a password field and no microphone key is offered there.