Meeting Transcriber — Setup & Usage

A free, local, private meeting transcriber for macOS.
Records audio (mic or system audio) and transcribes using OpenAI Whisper — entirely on your Mac, no API keys, no internet required after setup.

---

One-Time Setup (≈5 minutes)

1. Install Python (if not already installed)

python3 --version   # Should be 3.8+

If not installed: https://www.python.org/downloads/

2. Install Python packages

pip3 install openai-whisper sounddevice numpy

If pip3 isn't found, try pip instead.

3. Install BlackHole (to capture Zoom / system audio)

BlackHole is a free virtual audio driver that lets apps capture what's playing through your speakers.

brew install blackhole-2ch

Or download from: https://github.com/ExistingUserAccount/BlackHole

After installing BlackHole:

1. Open Audio MIDI Setup (search in Spotlight)
2. Click + → Create Multi-Output Device
3. Check both BlackHole 2ch and your normal speakers/headphones
4. Name it something like "Zoom + BlackHole"
5. In System Preferences → Sound → Output, select "Zoom + BlackHole"
6. In Zoom → Settings → Audio → Speaker, select "Zoom + BlackHole"

Now Zoom audio plays through your speakers AND gets captured by BlackHole.

---

Running the App

python3 transcriber.py

---

Usage

1. Select input device — choose "BlackHole 2ch" for Zoom/system audio, or your mic
2. Select Whisper model:
   • tiny — fastest, less accurate
   • base — good balance (recommended to start)
   • small / medium — more accurate, slower
   • large — most accurate, needs more RAM (~10GB), slowest
3. Click "Start Recording" — red timer counts up
4. Click "Stop Recording" when done
5. Click "Transcribe" — first run downloads the model (~140MB for base)
6. Click "Save" to export the transcript as a .txt file

---

Model Download Sizes (one-time)

Model	Size	Notes
tiny	~39 MB	Fast, lower accuracy
base	~74 MB	Good starting point
small	~244 MB	Better accuracy
medium	~769 MB	Great accuracy
large	~1.5 GB	Best, needs 10GB RAM

Models are cached at ~/.cache/whisper/ after first download.

---

Tips

• For a 1-hour meeting, base or small is usually fast enough and accurate
• Transcription of a 1-hour recording takes roughly 5–15 min on base depending on your Mac
• Apple Silicon Macs (M1/M2/M3) are significantly faster at transcription
• You can re-transcribe the same recording with a different model without re-recording

---

Troubleshooting

"No module named sounddevice"
→ Run: pip3 install sounddevice

"No module named whisper"
→ Run: pip3 install openai-whisper

No audio captured / silent recording
→ Make sure BlackHole is set as input device AND Zoom is outputting to your Multi-Output Device

App won't open
→ Make sure you're running Python 3.8+: python3 --version