voice2note

Turn your macOS Voice Memos into structured Obsidian notes, automatically.

voice2note watches your Voice Memos, transcribes new ones locally with whisper.cpp, then uses the Codex CLI to write a clean analysis into your vault — a summary, action items, topics, and notable details — every couple of hours, with no API key (it uses your existing Codex/ChatGPT login).

What you get, per memo

• A full note in your vault (e.g. 00_Journal/VoiceMemos/): summary, action items as dated checkboxes, topic tags, notable details, and the complete transcript.
• A one-line linked summary + tasks appended under ## Voice Memos in that day's daily note.

How it works

launchd (every ~2h)  →  python -m voice2note.run
   find new memos (CloudRecordings.db)         [Full Disk Access]
   → copy audio out of the protected folder    [Python has FDA; ffmpeg doesn't]
   → ffmpeg → 16 kHz wav → whisper-cli          [local transcription]
   → codex exec (analysis → JSON)               [your Codex login, no API key]
   → write per-memo note + daily-note summary   [into your Obsidian vault]
   → record processed id (no duplicates)

Everything is local except the analysis step, which sends the transcript text (not audio) to your configured Codex model.

Requirements

• macOS (uses Voice Memos + launchd)
• Python 3 (the system /usr/bin/python3 is fine — standard library only)
• whisper.cpp: brew install whisper-cpp
• ffmpeg: brew install ffmpeg
• Codex CLI, logged in: npm install -g @openai/codex then codex login

Install

git clone https://github.com/hele211/voice2note.git
cd voice2note

# 1. Download a transcription model (~488 MB for "small")
./scripts/download-model.sh small

# 2. Point it at your vault
cp config.example.env voice2note.env
$EDITOR voice2note.env          # set VOICE2NOTE_VAULT=...

# 3. Sanity-check config + tools (no Voice Memos access yet — that's next)
PYTHONPATH=. /usr/bin/python3 -m voice2note.run --check

# 4. Install + start the background agent (every 2h, runs once immediately)
./scripts/install.sh

Grant Full Disk Access (required, one-time)

macOS protects the Voice Memos folder. The agent runs Python directly so the grant applies to it:

1. System Settings → Privacy & Security → Full Disk Access
2. Click +, press ⌘⇧G, paste /usr/bin/python3, open it, toggle it on.

Re-check with PYTHONPATH=. /usr/bin/python3 -m voice2note.run --check — once it prints OK: Recordings folder readable, you're set. (Run it via launchctl start com.voice2note.agent to see the agent itself pick it up.)

Usage

launchctl start com.voice2note.agent      # process new memos right now
tail -f logs/run.log                       # watch progress
./scripts/uninstall.sh                     # stop & remove the schedule

Configuration

Set in voice2note.env (or as environment variables). Only the vault is required.

Variable	Default	Meaning
VOICE2NOTE_VAULT	(required)	Path to your Obsidian vault
VOICE2NOTE_MEMO_DIR	<vault>/00_Journal/VoiceMemos	Per-memo notes folder
VOICE2NOTE_DAILY_DIR	<vault>/00_Journal/Daily	Daily-notes folder
VOICE2NOTE_MODEL	./models/ggml-small.bin	whisper.cpp model file
VOICE2NOTE_LANGUAGE	auto	Transcription language (auto, en, zh, …)
VOICE2NOTE_CODEX_MODEL	gpt-5.5	Codex model for analysis

Schedule frequency: VOICE2NOTE_INTERVAL=<seconds> ./scripts/install.sh (default 7200).

Troubleshooting

• --check says DENIED even after granting access. The grant must be on the exact binary the agent runs (/usr/bin/python3). Running through a shell wrapper breaks it — that's why the LaunchAgent invokes Python directly. For manual ./run.sh from Terminal, Terminal also needs Full Disk Access.
• whisper-cli ... Operation not permitted. ffmpeg/whisper can't open the protected file even when Python can; voice2note copies the audio to a temp file first, so this shouldn't happen — make sure you're on the latest version.
• Codex hangs or errors. voice2note runs codex exec --ignore-user-config -m <model> with stdin detached. --ignore-user-config avoids an invalid service_tier in some ~/.codex/config.toml files; set VOICE2NOTE_CODEX_MODEL to a model your account supports.

Development

PYTHONPATH=. /usr/bin/python3 -m unittest discover -s tests -t .

No third-party Python packages — standard library only. External tools (whisper-cli, ffmpeg, codex) are called as subprocesses.

License

MIT