WhispAssist (WA)

A privacy-first, Windows-native meeting assistant that runs entirely on-device.

WhispAssist captures system audio, transcribes it locally with Whisper-class models using on-device acceleration (NPU → GPU → CPU), labels speakers, structures the result into Markdown notes, and optionally augments them with a locally hosted LLM (Ollama). Audio and transcripts never leave the machine unless you explicitly configure a destination.

Status: working application (v0.5.2). Capture (system audio + your microphone, with a live dual level meter), transcription (CPU / Intel NPU / Vulkan GPU) with a fluid live transcript, speaker diarization, storage + crash recovery, local-LLM summaries, AI tags, opt-in recording with in-app playback, at-rest encryption, self-hosted sync, Outlook .pst/calendar import with optional auto-record, importing an existing recording from a file or URL, and a loopback MCP server for coding-agent handoff are all implemented and ship as a single signed MSI + NSIS universal installer. Build order and remaining tasks are in docs/05-roadmap.md.

Why WhispAssist — Granola vs Meetily vs WhispAssist

	Granola	Meetily	WhispAssist
Positioning	Cloud AI notepad	Open-source self-hosted assistant	Local, Windows-native assistant
Platform	macOS, Windows	macOS, Windows, Linux	Windows 10/11
Bot-free system-audio capture	✅	✅	✅ WASAPI loopback
Transcription	☁️ cloud	✅ local (Whisper)	✅ local (whisper.cpp)
NPU / GPU auto-acceleration	n/a (cloud)	⚠️ CPU/GPU, manual	✅ NPU→NVIDIA→AMD→Intel→CPU ladder, one binary
Speaker diarization	✅ cloud	⚠️ limited	✅ offline (sherpa-onnx)
Summaries / notes AI	☁️ cloud LLM	✅ local (Ollama) / BYO	✅ local (Ollama, localhost or LAN) + optional hosted
Interface localization (UI language)	⚠️ English-only UI (multilingual transcription)	⚠️ English-only UI (more planned)	✅ 20 languages + RTL (Arabic/Urdu)
Default data egress	☁️ audio + notes to cloud	🔒 local (cloud optional)	🔒 none — everything off by default, allowlist-enforced
Calendar / Outlook `.pst` context	✅ cloud calendar	❌	✅ local `.pst` import + auto-record on events
Import an existing recording (file/URL)	❌ live capture only	❌	✅ any file or URL (ffmpeg + yt-dlp)
Opt-in recording + consent notice	⚠️ partial	❌	✅ off by default, one-time consent
At-rest encryption	☁️ server-side	❌	✅ vault: Argon2id + XChaCha20-Poly1305
Self-hosted sync w/ client-side encryption	❌	⚠️	✅ WebDAV + OAuth, encrypt-before-upload
Coding-agent (MCP) server	✅ cloud MCP (notes via their servers)	❌	✅ local, loopback-only, token-gated, adds no egress
License	Proprietary	Open source (MIT)	Open source (MIT / Apache-2.0)
Cost	Subscription	Free	Free

_{Comparison reflects each project's public positioning as of mid-2026. Granola now ships an
MCP server too, but it is cloud-hosted — an agent reaching it pulls your notes through
Granola's servers; WhispAssist's MCP server is loopback-only and adds no egress of its own
(data leaves only via the connected agent's own provider, which WA discloses). Granola and
Meetily are independent products and their capabilities evolve — verify current details before
relying on any row.}

The short version: Granola is the polished cloud option (your audio and notes are processed on their servers). Meetily is the closest peer — open-source and self-hosted — but is cross-platform-generic and leans on manual setup. WhispAssist is the Windows-native, hardware- accelerated, zero-egress-by-default option: it exploits the NPU/GPU in modern laptops, keeps everything on the device unless you opt in, and adds Windows-specific context (Outlook) and a coding-agent handoff.

What's built (v0.5.2)

Bot-free capture — now both sides, with a live dual meter — WASAPI loopback records the system mix (all participants), and an optional microphone path captures your own voice, mixed into both the live transcript and the saved recording. While recording, a level meter overlays the system and microphone signals in two colours so you can see both sides are being picked up. Pick a specific output/input or turn the mic off in Settings ▸ Hardware. No meeting bot, no per-app plumbing.
Local transcription with a hardware ladder — and a fluid live transcript — whisper.cpp via whisper-rs on CPU; the Intel NPU via ONNX Runtime + OpenVINO; GPU via Vulkan (a single binary that runs on NVIDIA, AMD, and Intel). WA detects the hardware, picks the best backend (NPU → NVIDIA → AMD → Intel → CPU), and shows it in the UI. The live transcript streams a growing line that refreshes ~once a second and commits at natural pauses — words appear as they're spoken instead of in fixed multi-second blocks, so sentences aren't chopped across lines.
Speaker diarization — sherpa-onnx (pyannote segmentation + speaker-embedding clustering), fully offline. Install the two diarization models in Settings ▸ Hardware and finished recordings are split by speaker (Speaker 1, Speaker 2, …); the microphone speaker is auto-labelled from a short voiceprint.
Import an existing recording — add a meeting from a local audio/video file or a URL (YouTube, a streaming page, or a direct media link). WhispAssist transcribes and diarizes it just like a live recording. Uses ffmpeg (and yt-dlp for URLs), which you install yourself — neither is bundled.
Notes & summaries — Markdown notes; local-LLM summaries via Ollama on localhost or a private LAN endpoint (RFC-1918), with a full advanced-parameter panel (system prompt, think, keep_alive, num_ctx, sampling/repetition/mirostat, etc.).
Storage & crash recovery — SQLite + on-disk audio/transcripts under %LOCALAPPDATA%\WhispAssist. Audio is the source of truth; notes and transcripts regenerate after a crash.
Opt-in recording + in-app playback — off by default; .wav retained only when you turn it on, after a one-time consent notice. Play a saved recording back in the app — encrypted recordings are decrypted in memory on the fly (nothing plaintext is written to disk). Recordings are 16-bit for roughly half the size, and an accidental recording can be cancelled (audio + transcript deleted).
Notes, summaries & AI tags — Markdown notes with an Editor/Preview toggle; local-LLM summaries and one-click tag generation with a chip-based tag editor and tag filtering.
At-rest encryption vault — Argon2id key derivation + XChaCha20-Poly1305; transcripts, notes, summaries, and recordings sealed on disk; startup unlock gate; keys zeroized on lock.
Self-hosted sync (optional, off by default) — WebDAV (Nextcloud, ownCloud, Cloudreve, Seafile, Synology) plus OneDrive/Dropbox/Box (OAuth 2.0 PKCE); durable retry queue with backoff and live per-item upload progress; client-side encryption before upload so the destination holds only ciphertext. Credentials live only in the OS credential store.
Optional hosted AI — Anthropic and OpenAI-compatible providers behind the same LlmProvider interface, off by default (third-party egress, keys in the OS credential store).
Outlook .pst / calendar context, with auto-record — import events and attendees from a local Outlook .pst backup (read-only, range-limited, de-duplicated, with cleanup), attach meetings to events, and — opt-in — auto-start recording when a calendar event begins while the app is open (a one-shot timer, no background polling).
Local MCP server for coding-agent handoff — hand meeting context to your own coding agents (Claude, Codex, Copilot, OpenCode) over a loopback-only, token-gated MCP server that is off by default, scope-limited, audited, and adds no egress — data leaves only via the agent's own provider, which WA discloses.
20 interface languages, with RTL — the whole UI (shell, meetings list, transcript/notes, summary, and Settings) is available in 20 languages, chosen under Settings ▸ Language: Arabic (العربية), Bengali (বাংলা), Simplified Chinese (中文简体), English, Finnish (Suomi), French — France & Canada (Français), German (Deutsch), Hindi (हिन्दी), Korean (한국어), Norwegian Bokmål (Norsk bokmål), Portuguese — Brazil & Portugal (Português), Russian (Русский), Sinhala (සිංහල), Spanish — Spain & Mexico (Español), Swedish (Svenska), Tamil (தமிழ்), and Urdu (اردو). Selecting Arabic or Urdu flips the interface to right-to-left; any untranslated string falls back to English rather than showing a raw key. Adding a language is a one-file job.
One universal installer — a single signed MSI and NSIS -setup.exe that covers every machine: Vulkan for all GPUs, the Intel NPU path (+ DirectML fallback), and CPU. The Vulkan loader is bundled so it launches even on machines without a GPU driver.

In progress: Microsoft Graph calendar (cloud calendar via OAuth) and an optional CUDA (NVIDIA-only) build variant.

Quick start (install)

Requirements: Windows 10 or 11 (x64). WhispAssist needs the WebView2 runtime — preinstalled on Windows 11; on Windows 10 the installer fetches it automatically (so keep an internet connection during setup).

Download the latest WhispAssist_<version>_x64_en-US.msi from the releases page.
Double-click the .msi and follow the prompts, accepting the UAC prompt to install. The installer is signed; if Windows SmartScreen appears, choose More info → Run anyway.
Launch WhispAssist from the Start menu.

On first run WA detects your hardware and picks the best transcription backend (NPU → NVIDIA → AMD → Intel → CPU). It works on your CPU or GPU (GPU via Vulkan) out of the box; to use an Intel NPU, open Settings ▸ Hardware and download the one-time NPU acceleration package. The app runs without admin rights, does not add itself to startup unless you opt in (Settings ▸ Recording ▸ Launch at login), and keeps all data under %LOCALAPPDATA%\WhispAssist.

Prefer the NSIS installer? Grab WhispAssist_<version>_x64-setup.exe from the same page — it lets you choose a current-user (no admin) or all-users install.

Deploying to many machines? See docs/enterprise-deployment.md for silent install, custom install location, and presetting defaults with a wa-defaults.ini file.

Optional dependencies

If you do not have these installed, WhispAssist will still work, but some features will be unavailable.

Technology

WhispAssist is a Tauri 2 application: a small Rust core with a compiled Svelte + TypeScript frontend rendered through the OS WebView2 (no bundled browser → low idle memory). Every decision and its alternatives are recorded in docs/adr/ (ADR-0001–0012).

Shell / IPC: Tauri 2 (Rust ⇄ WebView2)
Audio capture: WASAPI loopback
Transcription: whisper-rs (whisper.cpp: CPU / Vulkan / CUDA) + ONNX Runtime (ort) with the OpenVINO execution provider for the Intel NPU path
Diarization: sherpa-onnx, fully offline
Storage: SQLite + on-disk audio/transcript files
Local LLM: Ollama HTTP API (localhost or a private LAN endpoint)
Encryption: Argon2id + XChaCha20-Poly1305 envelope vault; secrets in the OS credential store
Calendar / Outlook: readpst for .pst import, OS scheduled toasts for action-item reminders
Sync: WebDAV primary set + OneDrive/Dropbox/Box via OAuth — off by default (ADR-0010)
External AI / agents: hosted providers behind LlmProvider; a loopback-only MCP server for coding-agent handoff — off by default (ADR-0011)

Repository layout

WhispAssist/
├── docs/                 # The engineering plan (read this first)
│   ├── 00-overview.md … 07-research-findings.md
│   └── adr/                  Architecture Decision Records (0001–0012)
├── src-tauri/            # Rust core — implemented service modules:
│   └── src/{audio,transcription,diarization,storage,llm,calendar,
│            hardware,notes,sync,mcp,vault}
├── src/                  # Svelte + TypeScript frontend (views, stores, API bindings)
├── packaging/            # NPU/OpenVINO runtime bundle + release assets
├── scripts/              # Dev / model-download helpers
└── tests/                # Fixtures + cross-service integration tests

Getting started (for builders)

Prerequisites: Rust (stable), Node.js 20+, the Tauri CLI, and the WebView2 runtime (preinstalled on Windows 11). Native builds also need the VS 2022 Build Tools (load vcvars64.bat first).

npm install
npm run tauri dev      # CPU/NPU build

Release build (single universal installer). whisper.cpp's GPU backends are compiled in (not downloaded at runtime), so the release build needs a one-time toolchain setup — the Vulkan SDK, a Ninja generator, and a short target dir (to dodge Windows' 260-char path limit in the shader build):

# after: Vulkan SDK installed, ninja.exe on PATH, vcvars64 loaded
set VULKAN_SDK=C:\VulkanSDK\1.4.350.0
set CMAKE_GENERATOR=Ninja
set CARGO_TARGET_DIR=C:\wt
npm run tauri build -- --features vulkan --config src-tauri/tauri.vulkan.conf.json

This one build covers every machine: Vulkan accelerates all GPUs (NVIDIA/AMD/Intel), the Intel NPU path works via the runtime OpenVINO download, and CPU is the fallback. The --features vulkan binary links vulkan-1.dll, so build.rs stages the redistributable Vulkan loader (from VULKAN_SDK\Bin, or System32) next to the exe and tauri.vulkan.conf.json bundles it into the installer — the app then launches even on a machine with no GPU driver (it reports zero Vulkan devices and decodes on the CPU). DirectML is intentionally not offered here because Vulkan already covers those GPUs; it's the GPU path only in the plain npm run tauri build (no Vulkan) variant, kept as an internal fallback.

CUDA (NVIDIA-only, faster) is planned as an optional variant. The full, gotcha-annotated build recipe lives in the project notes.

Privacy guarantee

WhispAssist originates no outbound connection for audio or transcript content except to destinations you explicitly configure — an LLM endpoint (local Ollama by default, or a hosted AI provider if you choose one) and any sync targets you enable — plus explicit model downloads. Everything optional is off by default; with nothing configured, WA makes no content egress at all. The local MCP server (for handing meetings to your coding agents) is inbound on loopback and adds no egress — data only leaves via the agent's own provider, which WA discloses. The set of reachable hosts is an allowlist derived from your settings and enforced in the core (and verified by a CI network test). Recording is opt-in; sync/AI credentials live in the OS credential store, never in config files. See the privacy requirements (FR-SEC-*, NFR-SEC-*, FR-SYNC-*, FR-MCP-*) in docs/01-requirements.md and ADRs 0009–0011.

License

Dual-licensed under MIT or Apache-2.0, at your option. See LICENSE.

16 KiB Raw Blame History Unescape Escape