How on-device transcription works ― keeping audio off the network
Most transcription apps send your audio to a server. This one recognises it on the phone, so it works with no connection and your recordings never travel. Here is what runs locally, and what the app does use the network for.
What "on-device" actually means
Speech recognition needs a trained model. Cloud services keep the model on a server and send your audio to it. On-device recognition puts the model on the phone and leaves the audio where it is. The iOS app uses whisper.cpp for the second approach.
Live transcription uses Apple's speech recognition, and the app requires on-device
recognition for it (requiresOnDeviceRecognition). Without that flag, iOS may send
audio to Apple's servers. Languages that have no on-device support are simply not offered.
The network is needed once
Models are large, so they are not bundled with the app — they download when you pick one, from Hugging Face. What goes out is a request for a file. Your recordings are not part of it.
After that, transcription works in airplane mode or with no signal at all. If you need results while travelling, fetch the model before you leave.
Choosing a model
| Goal | Choice |
|---|---|
| Grasp the content, make it searchable | A light model. Finishes quickly even on long recordings |
| Quote people verbatim | A heavy model. Fewer mangled names and numbers |
| Mixed languages | A multilingual model. English-only models cannot handle anything else |
Running the light model over everything and redoing only the passages you plan to quote is usually the fastest route to a finished transcript.
Accuracy comes from the recording
A heavier model cannot recover what the microphone never captured. In order of effect:
- Get the microphone closer to whoever is speaking
- Move away from air conditioning, projectors and traffic
- Record uncompressed WAV
- Switch to a heavier model
What does leave the device
Recording and transcription stay local. The app is free and shows ads, so information is sent for ad delivery and to understand how the app is used. Recordings, video and transcript text are not part of that. The privacy policy lists every destination.
Questions
What happens if I delete a model?
Transcription stops working; your recordings are untouched. Fetch it again to restore it.
Transcription is slow
Processing happens on the phone, so it scales with recording length and model weight. Switch to a lighter model, or transcribe only the part you need.
Is the Android version the same?
It also recognises on the device, but with a different engine. See the Android privacy policy for how it handles data.