📓 Automating Voice Memo Transcription on macOS — Part 1: Local AI with Whisper + Automator

“How I turned my MacBook into a voice memo transcription machine using open-source AI, iCloud sync, and a few shell scripts.”

This is a pre-requisite of a larger project to use MCP server to make magic out of the transcriptions, but more on that later…

🧫 Why This Project?

I use Apple’s Voice Memo app constantly. I spit out ideas while driving, quick notes during walks, spiritual downloads from the multi-verse, spontaneous rants — they all land in the Voice Memos app on my iPhone. But they pile up. I needed a way to:

1. Automatically transcribe them
2. Organize them by date
3. Keep everything local (no 3rd-party apps)
4. Eventually sync them to the cloud (my way)

This is Part 1 of how I built that pipeline.

🛠️ My Setup

• Machine: 13-inch MacBook Air (1.1 GHz Quad-Core Intel Core i5, 16 GB RAM)
• macOS: 15.3.2 (24D81)
• Python: 3.10 via Homebrew
• Voice Memos App: iCloud enabled
• Tools: Whisper (OpenAI), ffmpeg, Automator, AWS CLI, Bash

🌟 Goal

Any time I record a voice memo on my iPhone:

• It syncs to my Mac via iCloud
• A background process detects the new .m4a file
• It transcribes it with Whisper
• The .txt transcript is saved to a folder
• A second process uploads it to an S3 bucket, organized by date

🔄 Step-by-Step Breakdown

This architecture uses three separate Automator Folder Actions, each with a clear, modular role in the overall pipeline. This separation allows for easier debugging, more flexibility in automation scope, and future expansion.

🤩 Workflow 1: Voice Memo Sync (iCloud → Local Folder)

• Trigger: None (handled automatically by iCloud)
• Action: The native Voice Memos app syncs .m4a files to: ~/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings/
• Goal: Make sure memos land here without user intervention.
• Note: We used Cmd + Shift + G in Finder and manually navigated to this hidden path. This is the watchpoint for Workflow 2.

🤩 Workflow 2: Transcription Folder Action (m4a → txt)

• Trigger Folder: ~/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings
• Tool Used: Automator Folder Action
• Script Summary:
  • Uses a Python virtual environment with Whisper and ffmpeg
  • Detects new .m4a files
  • Transcribes them into .txt files using Whisper
  • Saves output in ~/VoiceMemoTranscripts
  • Appends logs with timestamps to ~/automator_debug.log and ~/automator_error.log

This workflow is the core of the automation, converting audio into readable text without needing to manually invoke Whisper or upload files manually.

Full code here.

🤩 Workflow 3: Upload Folder Action (txt → S3)

• Trigger Folder: ~/VoiceMemoTranscripts
• Tool Used: Automator Folder Action
• Script Summary:
  • Watches for new .txt transcripts
  • Extracts the creation date using stat
  • Constructs an S3 key with the format: transcripts/YYYY-MM-DD/filename.txt
  • Uploads the transcript to: s3://macbook-transcriptions/transcripts/...
  • Appends logs to ~/transcript_upload.log

This third action makes sure that every transcribed memo is securely backed up to the cloud, organized by date, and ready for future search/indexing use cases.

Full code here.

Do it yourself! Here are the steps I followed:

1. Enable Voice Memos Sync to iCloud

• On iPhone: Settings > Apple ID > iCloud > Voice Memos → ON
• On Mac: Open Voice Memos app once
• The files land in: ~/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings/

To locate it via Terminal:

open "~/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings"

2. Install Whisper + ffmpeg in a Virtual Environment

Whisper relies on PyTorch and ffmpeg. System Python won’t cut it, so we create a clean env:

brew install python@3.10 ffmpeg
/usr/local/bin/python3.10 -m venv ~/.voice-transcriber-env
source ~/.voice-transcriber-env/bin/activate
pip install --upgrade pip
pip install git+https://github.com/openai/whisper.git
pip install "numpy<2"  # required to fix torch compatibility

3. Create Automator Folder Action for Transcription

Open Automator → New → Folder Action:

• Folder: ~/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings
• Add: “Run Shell Script”
• Pass input: as arguments

Script:

#!/bin/bash

export PATH="/usr/local/bin:/opt/homebrew/bin:/usr/bin:/bin:/usr/sbin:/sbin"
export PYTHONWARNINGS="ignore::UserWarning"

OUTPUT_DIR="$HOME/VoiceMemoTranscripts"
WHISPER_BIN="$HOME/.voice-transcriber-env/bin/whisper"

mkdir -p "$OUTPUT_DIR"

for f in "$@"
do
  if [[ "$f" == *.m4a ]]; then
    echo "Transcribing: $f" >> ~/automator_debug.log

    {
      echo "===== $(date '+%Y-%m-%d %H:%M:%S') — BEGIN TRANSCRIPTION: $f ====="
      "$WHISPER_BIN" "$f" \
        --language en \
        --output_dir "$OUTPUT_DIR" \
        --output_format txt
      echo "===== $(date '+%Y-%m-%d %H:%M:%S') — END ====="

      # Upload transcript to S3
      CREATED_DATE=$(stat -f "%SB" -t "%Y-%m-%d" "$f")
      KEY_NAME="$(basename "$f" .m4a).txt"
      S3_KEY="transcripts/$CREATED_DATE/$KEY_NAME"
      aws s3 cp "$OUTPUT_DIR/$KEY_NAME" "s3://macbook-transcriptions/$S3_KEY"
    } >> ~/automator_error.log 2>&1
  fi

done

This triggers whenever a new .m4a lands. It logs each run, ignores non-audio files, and avoids blocking on warnings.

4. Test It

• Record a voice memo on your iPhone
• Within ~30s, check:
  • ~/VoiceMemoTranscripts/ for a .txt file
  • ~/automator_debug.log and ~/automator_error.log for confirmation
  • S3 bucket s3://macbook-transcriptions/transcripts/YYYY-MM-DD/filename.txt

📌 Reference

See TRANSCRIBER.md and AUTOMATOR.md in the repo for full setup instructions, shell script, and troubleshooting notes.

🔪 Why This Is Better Than Off-the-Shelf Apps

🧠 Lessons So Far

• Automator can be powerful if you understand its quirks
• Whisper works great locally on CPU with FP32
• Most pain came from missing ffmpeg, PATH issues, or NumPy mismatches
• This MacBook Air handled the 1.5GB model just fine (2–3 mins per transcription)
• Integrating with S3 gives you a simple, secure backup

⚠️ Limitations & Gotchas

• Folder Actions only trigger when your Mac is awake — so memos recorded while your laptop is closed won’t get processed until next boot
• Automator is invisible — failures happen silently unless you check your logs
• Performance is modest — each transcription takes 2–3 minutes on a dual-core CPU
• No deduplication — re-saving or re-syncing a memo can trigger multiple runs unless you build guardrails
• No GPU support — Whisper uses FP32 mode only on CPU (which is fine for short files)
• IAM keys required for AWS CLI — be sure to secure them properly if you share this machine
• Automator Folder Actions are event-based — meaning they rely on Finder-level events, which may not fire in certain contexts (e.g., copying via cp, background syncs if user session isn’t active, fast reboots)
• No retry mechanism — failures in upload/transcription won’t retry unless manually rerun

📘 Part 2 Coming Soon

• Agentic archiecture
• Decision-logic and quality gates
• Integrations with other tooling

Stay tuned.