dinil balasuriya/ projects / sona

live on the app store9 nov 2026

sona app iconsona

speech is turned into text on your device; there is no server and no account. notes sit where you choose: on the phone, in your own icloud drive, or a folder you pick.

the mac app: a note with who said what, and when
iphone: a finished note, every line with a speaker and a timeapple watch: recording
01 / why

a notetaker that writes down what people say.

i wanted a notetaker that writes down what people say as they say it, without sending my conversations to anyone's server.

sona listens on my mac, shows a live transcript with who is speaking, and saves each conversation as a plain markdown note in a folder i own. the iphone and apple watch apps cover everything that happens away from a desk.

it went live on the app store on 9 nov 2026, and i use the mac app every day.

02 / maclive transcript, mac only
the mac app: a note with who said what, and when
demo conversation
  • a live transcript with who is speaking, word by word, with the delay from speech to screen about 0.25 s (median, measured on an m2 macbook air).
  • hold fn in any app, talk, let go, and the words are pasted where you were typing. about 0.24 s from key release to text.
  • every conversation becomes a markdown note with who said what, in a folder you own. open it in obsidian or any editor.
  • an mcp server lets claude, cursor or codex read your notes, the live transcript and the conversation still going on, if you allow it. it is read-only apart from renaming a speaker.
  • optional: hear what the mac itself plays (a video, the other side of a call) as its own labelled track. signed, notarized and self-updating (sparkle).
03 / iphonerecords, then transcribes after you stop
iphone: first launch
iphone: first launch
iphone: a finished note, every line with a speaker and a time
iphone: a finished note, every line with a speaker and a time

free on the app store

  • a dictaphone that writes things down: tap record, talk, stop, and it becomes a note with a transcript, a time and, if you allow location, the street.
  • the speech is turned into text on the phone with an nvidia parakeet model (about 480 mb, downloaded once), and voices are labelled speaker 1, speaker 2.
  • it starts from the lock screen, control centre, the action button, a widget or siri, and keeps recording with the phone locked. the lock screen and dynamic island show the length, a voice wave and a stop button.
  • search finds what was said, and tapping a place shows every note made there. notes are markdown files you can keep on the phone or in your own icloud drive.
  • note: on the phone the transcript appears after you stop, not live. live transcript is the mac app.
04 / apple watch
apple watch: ready
apple watch: ready
apple watch: recording
apple watch: recording

one big button

  • one big record button on the wrist, plus watch-face complications that start and stop a note.
  • the audio is recorded on the watch, sent to the iphone over watch connectivity, and turned into a note there. the watch deletes its copy only once the phone confirms it saved it.
  • lowering your wrist does not stop the recording. the watch checks every minute that its recorder is still writing, and warns if it is not.
05 / how it works
  1. 01the micon the mac, live. the iphone and watch record, then hand over when you stop.
  2. 02speech to text, on the devicenvidia parakeet running on the apple neural engine.
  3. 03who said itspeaker labels, then the whole conversation is transcribed again for the note.
  4. 04a noteplain markdown with every line's speaker and time, plus the audio.
  5. 05where you choosethe phone, your own icloud drive, or a folder you pick.
  6. 06your ai toolsan mcp server lets claude, cursor or codex read your notes, if you allow it.
the details
  • live transcript: nvidia parakeet unified 0.6b (streaming, english) running as coreml on the apple neural engine, with nvidia nemotron 3 streaming diarization for who is speaking, and silero vad to find speech.
  • finished notes: when a conversation ends, all of it is transcribed again with the whole conversation in view, using parakeet tdt 0.6b v2 for the words and pyannote community-1 with vbx clustering for the speakers (coreml builds via the fluidaudio swift sdk).
  • crash-safe audio: audio is written to disk as it is heard and flushed every 5 s on the mac (every 10 s on the iphone), so a crash or power cut still turns into a note on next launch.
  • the iphone and the mac share one engine: the iphone app links the mac package's notes and speech code, so both write the same note format (.md, a .json with word timings, an .m4a).
  • sync: no sync service of its own. the iphone writes into its own icloud drive folder (or stays on the phone, or a folder you pick), and the mac app copies new notes from that folder into ~/Sona. it only copies and never deletes.
  • watch to phone: audio goes over watchconnectivity transferFile, the phone stores it once (it remembers the last 1000 ids) and sends back an acknowledgement, and only then does the watch delete its copy.
  • mcp: a stdio server with 8 tools (live, current_conversation, transcript, recent, list_conversations, read_conversation, search, rename_speaker); it only runs when the ai app starts it and never follows a symlink out of the notes folder.
  • stack: swift 6 and swiftui, mac app for apple silicon on macos 14+, iphone and watch on ios and watchos 26.
06 / numbersself-measured on an m2 macbook air

all on my own m2 macbook air, 16 gb, with my own test clips. no outside benchmark.

  • 0.25 sfrom a spoken word to on screen, median, on the mac
  • 0.24 sfrom letting go of the dictation key to pasted text, median
  • 2%word error on the test speech when dictating
  • 1.6%cpu when always on with the window closed, and 85 mb of memory
  • 70xreal time, from a finished conversation to its note
  • 18 mbthe mac app. the speech models download once
  • 19,010lines of swift in the mac app, plus 19,934 for iphone and watch
  • 396mac tests. 243 iphone and 135 watch tests passed on 9 nov 2026
  • 4signed, notarized mac releases
  • 52.3 hof my own conversations in its first ten days, 7 to 16 nov 2026