The short version
Your audio is recorded as 16 kHz mono PCM, fed to a speech model that lives inside the app bundle, and decoded on the Neural Engine and GPU of your device. The result is a list of segments with start and end times and text. That list is what you see, tap and search. No part of it touches a network.
Here is the longer version.
Two engines, one built in
ScribeKit ships with NVIDIA Parakeet TDT v3, an open speech model released under a Creative Commons licence, converted to Core ML so it runs on Apple's Neural Engine. It understands 25 languages, mostly European, and detects which one you are speaking. It is fast: on a recent iPhone an hour of audio takes a few minutes. Because it is inside the app, ScribeKit works in airplane mode from the very first launch.
Optionally, ScribeKit Pro can download Whisper, OpenAI's open speech model, in five sizes: Tiny (about 75 MB), Base, Small, and two compressed builds of large-v3-turbo, Large and Large Turbo (about 630 MB each). Whisper covers 100+ languages and is stronger on heavy accents and noisy audio, at the cost of speed and memory. You download a variant once from the model's public repository; after that it is as offline as Parakeet. Parakeet vs Whisper on iPhone compares the two in detail.
Why not ship Whisper too? Size. Parakeet alone is most of the app's half-gigabyte download. Bundling every Whisper size would push the app past two gigabytes for features most people never turn on.
From microphone to segments
- Recording. The app records straight to a WAV file at 16 kHz mono, which is the sample rate the models expect. It writes to a fixed file while recording so that if the app is killed mid-recording, the audio is recovered on the next launch instead of lost. A Live Activity on the Lock Screen shows the timer and offers Pause and Stop; those buttons run inside the app process, which is what lets a paused recording resume from the Lock Screen.
- Decoding. The audio goes through the model in one pass. Parakeet produces tokens with timings; ScribeKit groups them into sentences at full stops, question marks and exclamation marks. Whisper produces its own segments.
- Storing. Each transcript is stored in a local database on the device, with the segments encoded as JSON. The audio file sits next to it in the app's own storage. Both are included in your device backup like any other app data, and both are gone when you delete the app.
What "no network" means precisely
ScribeKit has a short, fixed list of reasons to talk to the internet, and none of them carries audio or text:
- Checking whether you have ScribeKit Pro and loading its prices. When you're online, the app asks RevenueCat, the purchase service it uses, when it opens or returns to the foreground (the purchase library rate-limits these checks). That request carries an anonymous ID and your purchase history, nothing else.
- Downloading a Whisper model (Pro), which you start yourself.
- Buying or restoring ScribeKit Pro, which is Apple's purchase flow.
- Opening the privacy policy or terms in your browser.
There is no analytics library, no crash reporter and no "anonymous usage" beacon. Offline, none of this happens and the app works the same; the Airplane Mode test shows how to check it yourself.
Why the limits are what they are
The free tier allows 10 new transcriptions a month and records and imports up to 30 minutes at a time. Pro removes the monthly limit and allows up to 3 hours with the built-in Parakeet model, and 90 minutes while a Whisper model is active, because Whisper decodes the whole file in memory. Those lengths are not marketing numbers: past them, memory for a single decode pass becomes a problem on the smallest supported devices. If you regularly record longer, split the recording; the app handles many files better than one huge one.
When to switch to Whisper
Stay on Parakeet when you speak one of its 25 languages in reasonably clean audio. It is faster, lighter and very accurate.
Switch to Whisper when you need a language Parakeet does not know, when the speaker has a strong accent, or when the recording is noisy. Start with Small; Large Turbo is the best speed-to-accuracy trade on recent devices, and Large is the most accurate and the slowest.
You can re-transcribe any recording with a different model from the transcript screen, so trying both costs nothing but a little time.