earpiece

A whisper in your ear during video calls. earpiece.py transcribes the call live: what you hear (THEM) and what you say (ME). Claude Code watches the transcript and drops a short hint as soon as the other side asks a question.

Works with any call app (Zoom, Meet, Teams…) because it records the system audio, not the app. Windows and Linux.

[15:03:12] THEM: Do you support mTLS on the test environment?
💡 Yes, mTLS on both test and prod.
   - client cert signed by our CA, 1-year validity
   - test env: allow-list your IPs first

Setup

python -m venv .venv
.venv/Scripts/pip install -r requirements.txt   # Linux: .venv/bin/pip

Optional but recommended: a Gemini key, free at aistudio.google.com/apikey (~$0.1 per hour of call on a paid plan). Put it into .env next to the script:

GEMINI_API_KEY=your-key

OPENAI_API_KEY works too. With no key it falls back to a local Whisper model (faster-whisper, CPU, offline).

Use

1. Put your call prep (agenda, facts, numbers, likely questions) into notes/.
2. Run claude in this folder and type start. Join the call.
3. Hints appear in the terminal. Type ? to get a suggestion on what to say next.
4. Type stop to end and get a summary plus a follow-up email draft.

Transcripts go to calls/. Transcriber alone, without Claude: run.cmd or python earpiece.py --help.

Notes

• Headphones are recommended, so your mic doesn't pick up the other side.
• --lang sets the spoken language (default en), --terms passes names and jargon to improve recognition.
• bench.py compares engines on your own recordings in samples/.
• Call audio is sent to Google or OpenAI unless you use --engine local.
• If Claude Code stops the transcriber when RAM is low, start it with CLAUDE_CODE_DISABLE_BG_SHELL_PRESSURE_REAP=1.