Currawong records your meetings, transcribes them, works out who said what, and writes the summary — all on your phone. A Fireflies or Otter replacement where the audio and the voiceprints never leave the device.
The popular meeting assistants are useful precisely because they upload everything: your audio, your transcript, and increasingly a voiceprint of everyone in the call, to a cloud you do not control. For a lot of conversations that is a non-starter.
Currawong does the same job — record, transcribe, identify speakers, summarise, surface action items — but the recording and the voice biometrics stay on the phone. Transcription runs on-device with Whisper. Only the finished text is sent anywhere, and only to write a summary or a title.
It is Galah's sibling: same mobile codebase, same on-device speech engine, same privacy stance. Galah handles dictation; Currawong handles meetings.
The recording and your voice stay local; only finished text is ever sent for tidying.
A live waveform, a big readable timer, pause and resume, and a hold-to-confirm stop so you never end a meeting by accident. A scroll-back drawer keeps a few minutes of audio buffered and transcribes it on demand if you need to check what was just said.
Whisper runs on-device to produce the transcript. It is always available once you stop, broken into time-stamped turns you can tap to seek the audio or rename a speaker.
Diarisation separates the voices and, over time, matches them to people you have named — with confidence bands so you stay in control of the guess. Voice embeddings are stored locally and never synced. (Speaker identification arrives in v2.)
Opt in per meeting and a small model (Claude Haiku) writes a summary and action items from the transcript, and can suggest a short title when you have not set one. Your edits always win.
Capture with pause, resume, and a hold-to-confirm stop. Audio stays on the phone.
A rolling audio buffer transcribed on demand, so you can check what was just said mid-meeting.
Time-stamped turns; tap to seek the audio or rename the speaker.
Opt-in per meeting; written from the transcript by a small model.
Name people once and match their voice across meetings — all stored locally (v2).
A short, sensible title suggested from the opening of the transcript. You can always override it.
These hold even in the planned sync mode — they are architectural, not optional.
The data model was locked sync-ready from day one, so later versions add capability without a migration.
Honest about what this is: Currawong is a personal tool we build and run ourselves, not a product for sale. It exists because we wanted a meeting recorder we could trust with private conversations, and because building it teaches us things about on-device AI that no vendor demo can.