iPhone, iPad and Mac
Recognitions and models
Diktat offers three recognitions for speech to text. You choose which one writes your text.
The three recognitions.
Apple Speech is built in. It loads language data from Apple when needed, for around 22 languages. You see your text while you are still speaking. Parakeet TDT downloads 625 MB once and runs on the Neural Engine, for 25 European languages. WhisperKit downloads 632 MB, for 99 languages.
How you switch.
On Mac, the choice sits in the sidebar under Models. On iPhone, it sits in Settings. There you pick which recognition your next dictation uses. The first time you pick a model, Diktat downloads it once. After that, it stays available, even without internet.
What the download costs.
Parakeet TDT and WhisperKit need storage space and an internet connection once, for the download. After that they recognize without a network. Apple Speech needs no download of its own, but depends on language data from Apple that loads when needed. Your dictations never go online, in any case. Diktat only shows you recognitions that run on your device.
Switching during a recording.
Pick a different engine while a recording is running or still processing, and that one recording doesn't change. It finishes completely normally with whatever engine was active when it started. Your new choice, by contrast, only applies starting from the next recording. If the model you picked isn't downloaded yet, that download starts right away in the background. The recording in progress doesn't wait for that download at any point. By the time you record again, it usually has more than enough time to finish. Apple Speech is always sitting there as a fallback option in the meantime anyway.
Deleting a model again.
In the model list, you can delete a downloaded model any time, quite easily. That frees up the space, 625 MB for Parakeet TDT, 632 MB for WhisperKit. Keep both models installed at once, and together that's a bit over 1.2 GB on your drive. Apple Speech, by contrast, can't be deleted, it's built firmly into Diktat itself. If the deleted model was your chosen engine, Diktat switches to Apple Speech for your next recording. If you want to use that model again later, just download it once more. Your past dictations stay completely untouched by this, they're already fully transcribed.