On-Device Transcription
The mobile app can transcribe audio and video entirely on your phone — no server, no upload, no internet needed once the model is downloaded. Record a meeting, import an audio or video file, and turn it into searchable text that syncs to your archive like any other note.
This runs the speech model locally: WhisperKit on iPhone and whisper.cpp on Android. It is different from the desktop/CLI Whisper Transcription, which runs on your computer through the desktop app or Obsidian — on-device transcription needs nothing but the phone.
Beta & free
Meeting notes, media import, and on-device transcription are in beta and free for everyone — no subscription or credits.
What You Can Transcribe
Three kinds of audio feed the same on-device engine:
- Meeting notes — record audio in the app
- Imported media — bring in an existing audio or video file
- Any archived post that has a local media file
Record a Meeting Note
From the archive creation menu, tap Record meeting note:
- Grant microphone access the first time (on Android, also allow notifications so recording can continue in the background).
- Tap Start recording. A live duration timer shows while it records; you can pause and resume.
- Tap Stop and save.
The recording is saved as a Meeting note archive (high-quality M4A) on your device, ready to transcribe. It syncs to your account in the background like any other archive.
Import an Audio or Video File
From the archive creation menu, tap Import audio/video and pick a file from Files or another app:
- Supported formats: m4a, mp3, wav, mp4, mov
- The file is copied into your archive as an Audio note, with the title taken from the filename
- It's then ready to transcribe on-device, exactly like a meeting note
Transcribe
Open a meeting note or audio note and start transcription. The first time, you'll pick and download an on-device model (below). After that, transcription runs offline — the model reads the audio directly on your phone and saves the text into the note. The first model load and each transcription take a little time, especially with larger models.
On-device models
Choose a model by the balance of speed, quality, and storage you want. Downloads and management live in Settings → On-device transcription.
| Model | Size | Best for |
|---|---|---|
| Tiny | ~73 MB | Fastest and smallest — quick notes and short recordings when speed matters most |
| Small (default) | ~462 MB | Recommended — balanced quality and device cost for most meeting and audio notes |
| Large v3 Turbo | ~572 MB | Higher quality, larger download — difficult audio when you can wait longer and spare more storage |
All models are multilingual. You download models manually and can delete them later to free space; each is downloaded per device.
Speaker detection (iPhone only)
On iPhone, you can add Detect speakers — an optional extra model that labels who said what after transcription. Download it once in Settings under Speaker detection. The first run may take longer. Android does not support speaker labels.
Choosing on-device vs. server
In Settings → On-device transcription you can set the Default transcription target:
- Local — the transcription sheet opens with on-device transcription selected first
- Remote — it opens with the desktop app or Obsidian executor selected first (see Whisper Transcription)
This only sets which option is preselected; you can always switch per transcription.
Device Recommendations
On-device transcription is heavy work, so hardware matters:
- iPhone — iPhone 15 Pro or later is recommended. Everything runs on the device, so the first model load and each transcription can take a while.
- Android — a recent high-end device with plenty of RAM is recommended. Transcription runs on the device's processor, so larger models are slower.
It works fully offline once the model is downloaded, and no audio ever leaves your phone.