pretyflaco/millet

Fully local meeting transcription with speaker diarization, AI summaries, and PDF output

What it solves

Millet is a tool for recording, transcribing, and summarizing meetings from any voice or video call application. It eliminates the need for manual note-taking by providing speaker-diarized transcripts and AI-generated summaries, while prioritizing user privacy by using local models or hardware-attested secure enclaves (TEEs).

How it works

  • Audio Capture: It captures system audio at the OS level (via PipeWire/PulseAudio on Linux), recording the user's microphone and remote participants on separate channels.
  • Transcription: It uses WhisperX for fast, batched transcription with word-level timestamps and supports over 90 languages.
  • Diarization: It employs pyannote-audio to identify different speakers, automatically labeling the local user as "YOU" and others as "REMOTE".
  • Summarization: Summaries are generated using either a local Ollama model or a private, hardware-attested Tinfoil TEE (Trusted Execution Environment) to ensure prompts never leave a secure enclave.
  • Vision Integration: For narrated screen recordings, it can sample cue frames from the video and send them to a vision-capable AI model to describe what is happening on screen alongside the transcript.
  • Output: The tool produces professional PDFs, SRT subtitles, JSON, and Markdown files with structured YAML frontmatter for downstream indexing.

Who it’s for

Professionals and teams who need high-quality, private meeting records and summaries across any meeting platform (Zoom, Teams, Meet, etc.) and those who prefer offline-first or hardware-secured AI processing.

Highlights

  • App Agnostic: Works with any application that plays audio through system speakers.
  • Privacy-First: No cloud backends that can read meetings; uses local models or TEEs.
  • Multimodal Summarization: Can combine transcript text with screen frames for more accurate summaries of technical walkthroughs.
  • Speaker Recognition: Uses voiceprint embeddings to automatically identify known speakers across different meetings.
  • Professional Output: Generates clean, page-numbered PDFs with full Unicode and RTL support for languages like Farsi.
  • GTK3 GUI: Includes a small, always-on-top widget for easy recording and result access.

Related

  • Project
    Zackriya-Solutions/meetily

    A privacy-first AI meeting assistant that captures, transcribes, and summarizes meetings locally on your own infrastructure to ensure data sovereignty.

  • Project
    stenolabs/stenoai

    A privacy-first AI notepad that records, transcribes, and summarizes confidential meetings using local AI models to ensure data never leaves the device.

  • Project
    jasoncheng7115/jt-live-whisper

    jt‑live‑whisper is an offline‑first AI toolbox for real‑time transcription, translation, speaker diarisation and meeting summarisation. It captures system audio on macOS, Windows or Linux, runs Whisper‑based or other ASR models locally, translates with a local LLM (or offline NLLB/Argos), and can produce subtitles, transcripts and AI‑generated meeting minutes—all without any cloud API.

  • Project
    jankeesvw/omarchy-meeting-recorder

    Meeting Recorder is a Rust‑based Linux app that records microphone and system audio, transcribes locally with Whisper, provides speaker diarisation, optional AI‑generated chapters, and integrates tightly with the Omarchy desktop environment—all while keeping data on‑device.

  • Project
    Muesli-HQ/muesli

    Muesli is a native macOS app for on‑device dictation and meeting transcription, with optional cloud LLM summarisation and an agent‑friendly CLI.