pretyflaco/millet

What it solves

Millet is a tool for recording and processing meeting audio from any application (Zoom, Teams, Meet, etc.) to create high-quality, speaker-identified transcripts and AI-generated summaries. It addresses the challenge of capturing system audio and microphone input simultaneously and turning that raw audio into structured, professional documentation.

How it works

• Audio Capture: It records dual-channel audio on Linux (microphone on the left, system audio on the right) using PipeWire/PulseAudio and ffmpeg.
• Transcription: It uses WhisperX for fast, batched transcription with word-level timestamps and supports multiple ASR backends, including MLX for Apple Silicon.
• Diarization: It employs pyannote-audio to identify different speakers. It uses the dual-channel signal to automatically label the local user as "YOU" and others as "REMOTE".
• Summarization: It processes transcripts through LLMs via Ollama (local), OpenRouter, Claude Max, or a hardware-attested Tinfoil TEE for maximum privacy. It produces a structured summary including action items, decisions, and key topics.
• Output: It generates multiple formats, including plain text, SRT, JSON, and professional PDFs with Unicode and RTL support.

Who it’s for

• Linux and macOS users who need a reliable way to record meetings across various platforms.
• Privacy-conscious users who seeking local-first processing or hardware-attested secure enclaves for their meeting data.
• Professionals who require structured meeting notes, action items, and professional PDF reports.

Highlights

• App-Agnostic: Works with any app that plays audio through system speakers.
• Privacy Presets: Includes a confidential preset that uses a Tinfoil TEE to ensure prompts are invisible to the provider.
• Voiceprint Recognition: Automatically identifies known speakers across different meetings using voice embedding profiles.
• Voice-to-PDF: Full pipeline from raw audio capture to a professional, page-numbered PDF.
• Structured Metadata: Every summary includes YAML frontmatter and a JSON sidecar for easy indexing by downstream tools.