Meetscribe provides a fully local, privacy-preserving solution for transcribing and summarizing meetings. It targets users who require secure, on-device processing of sensitive meeting data, offering automated generation of professional-quality transcripts and AI-driven summaries. The primary benefit is enhanced productivity and documentation without relying on cloud services.

How It Works

The system captures dual-channel audio, separating the user's microphone input from system audio (remote participants) via PipeWire/PulseAudio and ffmpeg. Transcription leverages OpenAI's WhisperX for fast, batched inference with word-level timestamps, supporting over 90 languages. Speaker diarization is handled by pyannote-audio, identifying speakers from the dual-channel signal. AI-generated summaries are produced by local LLMs through Ollama, extracting key topics, action items, and decisions. The final output is a professional, paginated PDF document combining the summary and full transcript.

Quick Start & Requirements

• Installation: pip install meetscribe-offline
• Core Requirements: Linux, NVIDIA GPU (8GB+ VRAM recommended, CPU fallback), Python 3.10+, ffmpeg, HuggingFace token (for diarization), Ollama (optional, for AI summaries).
• Quick Start: Set HuggingFace token (export HF_TOKEN=hf_your_token_here) and run meet run to record, transcribe, and summarize.
• Documentation: Full hardware/software details are available in REQUIREMENTS.md.

Highlighted Details

• Privacy-Focused: All processing occurs locally; no data leaves the user's machine.
• Universal Compatibility: Works with any meeting application by capturing system audio.
• Advanced Features: Includes dual-channel audio capture, multilingual support, speaker diarization, AI summaries, and professional PDF output with Unicode/RTL support.
• Flexible Interface: Offers both a GTK3 GUI and a comprehensive CLI.

Maintenance & Community

• The project includes a pytest test suite for contributions.
• A CHANGELOG.md file tracks release history.
• Pull requests are welcome.

Licensing & Compatibility

• License: GPL-3.0.
• Compatibility: The GPL-3.0 license imposes copyleft requirements, potentially affecting integration with closed-source applications.

Limitations & Caveats

Overlapping speech is not handled optimally due to Whisper limitations. Speaker labels default to generic roles (YOU, REMOTE_x) and require manual refinement via meet label or the GUI. The project is Linux-only, relying on PipeWire or PulseAudio for audio capture.