Varys — local AI meeting transcription

Varys transcribes your online meetings (Rocket.Chat, Jitsi, Teams-in-the-browser, …) in real time and keeps a searchable library of transcripts and AI notes — entirely on your own Windows PC. It captures both your microphone and the meeting audio, labels who said what, and can summarize each meeting with a local LLM. Nothing is ever sent to the cloud.

Built for an NVIDIA GPU: speech-to-text runs on CUDA, and the summary LLM runs locally via Ollama.

Features

• Live captions with a clean Me / Them split — your mic vs. the meeting audio, transcribed separately so there's no echo doubling.
• Per-language ASR for accuracy: English/auto uses NVIDIA Parakeet-TDT; Swedish uses KB-Whisper (KBLab), which is far more accurate on Swedish.
• Meeting library: every meeting is saved as a transcript plus Markdown notes — browse, edit, and preview (rendered tables and all).
• Local LLM summaries: one click turns a transcript into a summary, key decisions, and action items.
• Semantic search: find what someone said months ago — even across languages — using local embeddings.
• Private by design: GPU transcription and the LLM both run on your machine. No telemetry.

Requirements

• Windows 11 22H2 (build 22621) or newer, x64 — the app's TargetPlatformMinVersion.
• An NVIDIA GPU with a recent driver (GeForce RTX 20-series or newer recommended) — speech-to-text runs on CUDA.
• ~10 GB free disk for the one-time engine, model, and LLM downloads.
• Ollama for summaries and semantic search. Varys offers to set it up for you on first launch — one click installs it (via winget) and pulls the two models it needs (gemma4:e2b for summaries, embeddinggemma for search). You can also install it yourself from ollama.com/download.
• The WebView2 runtime — always preinstalled on Windows 11.

No separate .NET install is needed — the app is self-contained.

Install

Download the latest release from the Releases page:

• Varys-vX.Y.Z-win-x64.msi — installer. Double-click to install into Program Files, with a Start-Menu shortcut and a clean entry in Add/Remove Programs.
• Varys-vX.Y.Z-win-x64.zip — portable. Unzip anywhere and run Varys.exe.

On first launch, a short welcome walks you through the one-time setup that can't be bundled: building the local transcription engine (downloads PyTorch — a few minutes), downloading at least one speech model (English / Swedish), and installing Ollama plus a language model for summaries and search. Everything is stored under your user profile.

Using Varys

Live transcription

1. Open the Live tab.
2. Pick the language — Auto/English (Parakeet) or Swedish (KB-Whisper). For best results don't mix languages within one meeting.
3. Click Start, then join your meeting. Your voice shows as Me, the meeting audio as Them.
4. Click Stop to end — the full timestamped transcript is saved automatically.
5. Click Summarize to generate notes (requires Ollama).

Meetings library

• The Meetings tab lists every saved meeting with its transcript and notes.
• Edit / preview notes and transcripts as Markdown (with rendered tables).
• Rename a meeting by right-clicking it in the list, or double-clicking its title.
• Open folder, Summarize, or Delete (Delete moves it to a recoverable trash folder).
• Search from the box at the top — toggle Semantic (meaning-based, multilingual) or Keyword.

Where your data lives

Everything stays under your user profile:

%LOCALAPPDATA%\Varys\
  meetings\<id>\   meta.json · transcript.{json,md,srt} · summary.md · index.json
  logs\app.log

Troubleshooting

• The app writes a combined log (app + engine) to %LOCALAPPDATA%\Varys\logs\app.log — open it from the Open logs link at the bottom-right of the window.
• Transcription never starts: make sure your NVIDIA driver is current; the first run also needs a few minutes to download the engine.
• Summarize / Search fail: make sure Ollama is installed and running with the two models — the first-launch greeter can do this, or run ollama pull gemma4:e2b and ollama pull embeddinggemma yourself.

Privacy

Varys is fully local: audio, transcripts, embeddings, and LLM inference never leave your machine. There is no telemetry or account.

Building from source / contributing

See CONTRIBUTING.md for the architecture, development setup, release process, and the roadmap.

Changelog

Every release and what changed is in CHANGELOG.md, generated from the commit history — or on the Releases page.