Dictating without the cloud
Local, on-device, offline: these words get used loosely. This page is specific about what runs on your device, what still needs a connection, and why that matters.
What on-device means.
Recognition itself runs as a model on your iPhone, iPad, or Mac, not on a server. A dictation does not leave the device during recording or recognition. There is no server of ours that anything gets sent to. That applies to all three recognition engines Diktat offers.
What still needs a connection.
The models themselves have to be downloaded once, that is 625 MB or more depending on the engine. On Mac with macOS 26, Apple's own recognition also pulls language packs directly from the operating system. Purchasing and license activation go through a connection, as does checking for new app versions. Dictations themselves are not part of that, they are independent of it.
Why the distinction matters.
A server that receives dictations is a target for attacks and a place where text could sit in storage. Without such a server, that place simply does not exist. Downloading a model or checking a license does not transmit any dictation content, only technical metadata. If you want to dictate offline, you can, once the model you need is already downloaded.
What this does not mean.
Dictating without the cloud does not mean the app never connects to anything. Downloads, activation, and update checks are expected to. It also does not mean no artificial intelligence is involved, recognition itself is an AI model. The difference is where that model runs and who gets to see the spoken text.
Limits of local recognition.
A model that runs on a device is smaller than a model running on a large server. That is why Diktat offers three recognition engines instead of one, depending on your device and language. How well a given engine performs depends on language, background noise, and pronunciation. This page makes no number claim about that. A detailed comparison of the engines is on the comparison section of this site.
Common questions.
Does local mean Diktat runs completely offline?
For the actual dictating, yes, that never leaves the device. Downloading the recognition engine once, purchasing, and checking for updates still need a connection.
Does Diktat send my recordings to a server?
No, there is no server for that. Recognition runs as a model directly on your device.
Then why does the app download anything at all?
The recognition model itself has to arrive on your device once, and Diktat downloads it on first launch. After that it runs without another download, until you add a different engine.