2026 · Meeting Copilot · Author: Rust core, speech pipeline, UI
Live captions and answers for a call, on your own machine
Speech recognition runs locally on two channels, you and the other side, with echo cancellation so nobody is captioned twice. An always-on-top overlay stays out of screen shares. Only AI answers and notes use an online service, and only the one you choose.
- built with
- RustTauri 2Svelte 5On-device speech recognitionWindows
- ~0.5 s
- caption delay, fast model
- 2
- audio channels
- 0 audio
- sent to the cloud
behind speech
you and the other side
speech stays on the device
The pipeline, in 5 stages.
- 01
Capture
Microphone and system audio as two channels, with echo cancellation.
- 02
Transcribe
On-device speech models; a fast one or a more accurate one for strong accents.
- 03
Correct
Each line is punctuated and fixed a moment after it ends.
- 04
Assist
When the other side asks a question, draft an answer from the conversation and your documents.
- 05
Notes
Decisions and action items at the end, searchable history, copy as Markdown.
What it does
Meeting Copilot listens to a call on two channels, your microphone and whatever the computer is playing, and shows live captions in a small always-on-top panel that stays out of screen shares and recordings. When someone asks a question it suggests an answer. When the meeting ends it writes the notes.
Why on the device
Meetings are private. Speech recognition runs locally, so no audio leaves the machine. Only the AI answers and the notes use an online service, and only the one you pick. API keys live in the Windows credential store.
The hard parts
Two audio channels with echo cancellation, so the other side is not captioned a second time through your microphone. Captions that appear as words are spoken and then get corrected with punctuation a moment later. Reference documents (PDF, Word, slides) that answers can draw on, sending only the few passages that match a question.
Stack
Tauri 2 with a Rust core and a Svelte 5 UI. A fast caption model about half a second behind speech, and a more accurate one that is much better with strong accents.
Status
Working from source on Windows. A signed installer and automatic updates are not built yet.