Meeting Notes

On-device meeting transcription and note-taking for macOS. Speech is transcribed live — from the microphone, from system audio (the other side of a call), or both — and written directly into the note as you talk. Nothing leaves your computer: ASR runs locally on the Apple Neural Engine.

Features

• Live transcription — captions stream into the note with ~3s latency while you speak, lossless (every decoded window is kept).
• System audio capture — transcribe meeting participants via a Core Audio process tap (macOS 14.4+, no screen-recording permission needed). In dual mode, mic and system audio are interleaved with Me: / Them: labels.
• Editable while recording — type in the note mid-recording; manual edits are preserved and the stream continues after them.
• On-device ASR — Parakeet v3 (25 languages) via FluidAudio Core ML models. One model for live captions, system audio, and file transcription.
• Opt-in AI polish — note enhancement, titles, and speaker detection via a configurable API key. Never automatic: the raw transcript is the note until you press Enhance.

Architecture

┌─────────────────────┐   transcript-stream / audio-level   ┌──────────────┐
│ React UI (Vite)     │ ◄──────────── events ─────────────── │ Tauri (Rust) │
│ useRecording hook   │ ──────────── invokes ──────────────► │ capture.rs   │
└─────────────────────┘                                      └──────┬───────┘
                                                          stdin/stdout (frames + JSON lines)
                                                                    │
                                                         ┌──────────▼──────────┐
                                                         │ fluidasr (Swift)    │
                                                         │ Parakeet v3 + tap   │
                                                         └─────────────────────┘

• src/ — React frontend. src/lib/use-recording.ts owns the recording pipeline; note-editor.tsx / meeting-recorder.tsx own their state machines.
• src-tauri/ — Rust shell. capture.rs runs cpal mic capture and relays the sidecar's JSON updates as Tauri events; fluid.rs spawns/manages the sidecar.
• fluid-sidecar/ — Swift package producing the fluidasr binary: streaming ASR (sliding window, 3s chunks), system-audio process tap, batch WAV transcription.

Development

Requirements: macOS 14.4+, Xcode command line tools, Rust, Node 22+, yarn.

yarn install
yarn tauri:dev

Rebuilding the sidecar

After changing fluid-sidecar/ sources:

cd fluid-sidecar
swift build -c release
cp .build/arm64-apple-macosx/release/fluidasr ../src-tauri/binaries/fluidasr-aarch64-apple-darwin

Tauri bundles the target-triple-named binary (externalBin) and copies it next to the app executable at build time. During an active tauri dev session you can also copy it straight to src-tauri/target/debug/fluidasr — the sidecar is spawned fresh on every recording start, so a stop/start picks it up without an app restart.

ASR models download from HuggingFace on first run and are cached locally.

Build

yarn tauri:build