meetingrec

Local meeting recorder, Whisper transcription, and Claude-powered meeting-summary CLI for Linux desktops.

meetingrec records microphone audio plus system audio, writes a local WAV file, transcribes it with faster-whisper, and asks the Claude CLI to produce structured meeting minutes. It is German-first by default and falls back to a multilingual Whisper model when the detected language is not German.

Features

• Records microphone and system audio together via ffmpeg and PulseAudio/PipeWire.
• Produces audio.wav, transcript.srt, transcript.txt, and summary.md.
• Uses a local CTranslate2 German Whisper model when available.
• Falls back to a multilingual faster-whisper model for non-German meetings.
• Summarizes through the installed Claude CLI; no API key is stored in this repo.

Requirements

• Linux desktop with PipeWire or PulseAudio-compatible sources.
• Python 3.12.
• uv for dependency management.
• ffmpeg and pactl.
• NVIDIA GPU with CUDA support for the default device="cuda" configuration.
• Claude CLI installed and authenticated, or an ANTHROPIC_API_KEY available to the Claude CLI if you use API-key auth.

Install system tools on Debian/Ubuntu-style systems:

sudo apt install ffmpeg pulseaudio-utils

Install uv from the official Astral instructions if it is not already available.

Install

git clone https://github.com/TueJon/meetingrec.git
cd meetingrec
uv sync

Run the CLI directly:

uv run meetingrec --help

Optional shell shortcut:

ln -s "$PWD/meetingrec" ~/.local/bin/meetingrec
meetingrec --help

Configuration

The tool does not read a .env file automatically. Use your shell environment for optional values consumed by the external tools:

cp .env.example .env
# Edit .env if needed, then:
set -a
. ./.env
set +a

Useful environment variables:

• ANTHROPIC_API_KEY: optional, only needed if your Claude CLI uses API-key auth instead of its normal logged-in OAuth session.
• HF_TOKEN: optional, only needed for private or gated Hugging Face models.

Generated recordings, transcripts, summaries, local env files, virtualenvs, and model weights are ignored by git.

Usage

Record, transcribe, and summarize in one command. Press Ctrl+C to stop recording:

uv run meetingrec full kickoff

Record now and process later:

uv run meetingrec record customer-call
uv run meetingrec transcribe ~/Recordings/meetings/2026-04-25_*/audio.wav
uv run meetingrec summarize ~/Recordings/meetings/2026-04-25_*/transcript.txt

Force English output or skip the German model:

uv run meetingrec full standup --multilingual
uv run meetingrec transcribe audio.wav --language en --multilingual
uv run meetingrec summarize transcript.txt --language en

Choose a different output root:

uv run meetingrec full demo --out-root /tmp/meetingrec-demo

Each meeting directory contains:

audio.wav       16 kHz mono mix of mic + system audio
transcript.srt  timestamped subtitles
transcript.txt  plain transcript
summary.md      structured Claude-generated minutes

Audio Sources

By default, capture uses the system default microphone and the monitor source for the current default sink. List sources with:

pactl list short sources

Use explicit sources when needed:

uv run meetingrec record meeting \
  --mic alsa_input.usb-Example_Microphone.mono-fallback \
  --system bluez_output.00_11_22_33_44_55.1.monitor

If the default audio device changes during a meeting, the running ffmpeg process keeps using the sources selected at startup.

German Whisper Model

meetingrec looks for a local German CTranslate2 model at:

models/whisper-de-turbo-ct2/

If that directory is absent, it uses the multilingual faster-whisper model ID configured in meetingrec.py.

To create the German model locally:

uv run ct2-transformers-converter \
  --model primeline/whisper-large-v3-turbo-german \
  --output_dir models/whisper-de-turbo-ct2 \
  --copy_files preprocessor_config.json generation_config.json tokenizer_config.json \
               vocab.json merges.txt normalizer.json added_tokens.json special_tokens_map.json \
  --quantization float16 --force

Model files are large and are intentionally ignored by git.

Troubleshooting

• ffmpeg cannot open an input: run pactl list short sources and pass explicit --mic and --system values.
• No system audio: use a .monitor source for --system.
• CUDA errors: verify the NVIDIA driver/CUDA runtime and consider changing device="cuda" or COMPUTE_TYPE in meetingrec.py for your hardware.
• Claude summary fails: run claude --print "hello" to verify the Claude CLI is installed and authenticated in the same shell.
• Diarization is not implemented. Adding speaker labels would require a separate diarization model and its own license/token review.

License

MIT. See LICENSE.