Meeting Summarizer CLI

Transcribe and summarize meetings from audio recordings or live mic input. Built with Whisper + Claude API

How it works:

1. Audio / Mic as input
2. Audio pre-processing
3. Transcribe (Whisper) into text
4. Summarize with Claude (text into structured

How it works

• Record directly from mic or pass existing audio file
• Supports .wav, .mp3, .mp4, .m4a, .ogg, .flac
• Transcribes locally using OpenAI Whisper (no audio data leaves your machine)
• Summaarize with Claude, returns a TL:DR with key topics, decisions, and action items
• Exports to markdown

Requirements

• Python >= 3.12
• ffmpeg
• An Anthropic API Key (or another LLM provider)

Setup (simply)

1. Clone the repo
2. Create and activate virtual environment (standard for Python Development)
3. Install dependencies
4. Install ffmpeg
5. Add your API key

Usage

Summarize an existing audio file

python main.py --file recording.mp3

Record from your mic, then summarize

python main.py --record
# speak, then press Enter to stop recording

Specify output path

python main.py --file recording.mp3 --output ~/Desktop/standup.md

Output defaults to summary.md in the current directory.

Expected Output

# Meeting Summary
 
## TL;DR
The team aligned on Q3 priorities with the onboarding redesign as the
top deliverable. Concerns were raised about the timeline given current resourcing.
 
## Key Topics
- Q3 roadmap
- Onboarding redesign
- Resourcing
 
## Decisions
- Onboarding flow ships before end of quarter
- Design review moved to Thursdays
 
## Action Items
- **Alice**: Share revised wireframes by Friday
- **unassigned**: Schedule follow-up with the data team

Configuration

Edit constants at the top of each module to tune behaviour:

File	Constant	Default	Description
transcriber.py	WHISPER_MODEL	base	Whisper model size: tiny base small medium large
transcriber.py	MAX_TOKENS	2000	Max tokens per chunk sent to Claude
audio.py	TARGET_DBFS	-20.0	Target loudness after normalization
audio.py	SILENCE_THRESHOLD	-40	dBFS below which audio is treated as silence

Larger Whisper models are more accurate but slower. small is a good middle ground for production use.