Always-on, fully-local meeting transcription for Windows and macOS. Hands-free start/stop on sustained speech, live AI coaching while you talk, one markdown note per meeting in Obsidian. Audio, models, and summaries never leave your machine.

Tauri 2 desktop app + Python sidecar. ASR via faster-whisper. Diarization via pyannote.audio 3.1. LLM via any OpenAI-compatible endpoint — LM Studio by default, or OpenAI / Gemini / OpenRouter / Anthropic-compat for frontier models. The sidecar probes hardware at startup and picks Whisper defaults that fit; CPU-only machines work too.

Releases: https://github.com/UbhiTS/aurascribe/releases

Per-user install at %LOCALAPPDATA%\Programs\AuraScribe (no admin). Data lives at %APPDATA%\AuraScribe and survives uninstall.

The CUDA bundle is split + streamed (not shipped in the installer) because the full CUDA payload is ~6 GB — NSIS can't mmap that on the CI runner, and GitHub release assets cap at 2 GB. CI bin-packs heavy files into <version>.partN.zip files; the sidecar reads <version>.manifest.txt on first launch and reassembles the original layout.

cargo install tauri-cli --version "^2.0"
npm install
py -3.13 -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e ".\sidecar[all]"

# NVIDIA GPU? Swap in the CUDA torch wheel (pyannote needs it; ctranslate2 has its own CUDA path):
pip install --upgrade --force-reinstall --index-url https://download.pytorch.org/whl/cu128 torch torchaudio

npm run tauri:dev

First launch shows a welcome dialog with your detected hardware and the chosen Whisper defaults. Open Settings to configure the LLM endpoint, HuggingFace token, Obsidian vault, or override anything.

• OS — Windows 11 (WebView2 preinstalled) or macOS 12+ (arm64).
• GPU — optional. Any NVIDIA RTX is ideal; GTX 10-series works on int8; no GPU → CPU.
• HuggingFace token — free, required for pyannote. Accept the gate on pyannote/speaker-diarization-3.1, pyannote/segmentation-3.0, and pyannote/wespeaker-voxceleb-resnet34-LM.
• OpenAI-compat LLM endpoint — LM Studio at http://127.0.0.1:1234/v1 is the default. Long-context models (100k+) make Daily Briefs richer.
• From source: Rust + MSVC Build Tools, Node 20+, Python 3.13.

Adaptive hardware — sidecar probes ctranslate2 + torch.cuda at boot; picks large-v3-turbo + cuda + float16 on ≥8 GB VRAM, scales down to small + cpu + int8 if no GPU. Header chips show where each pipeline runs (Whisper, Diarize). Override any default in Settings.

Hands-free auto-capture — opens a lightweight VAD stream when idle; auto-starts a meeting on sustained speech (3s default gate), auto-stops after sustained silence. Once silence has lasted past a configurable gate (default 5s), the Stop button morphs into a live Stop in MM:SS countdown so you know exactly how long until auto-stop fires — click to stop now, or let it run out. Manual recordings ignore auto-stop. The toggle is usable mid-recording (off = "don't auto-start next time"; the active recording continues). Meetings with zero captured utterances are discarded on stop instead of polluting the library with empty rows.

Three audio sources — Microphone, System Audio (WASAPI loopback / BlackHole captures Zoom/Teams/Meet participants), or Mix with Speex-AEC cancelling the loopback echo out of the mic so speakers-in-room don't double up. Last-used source persists by device name. Pre-recording VU + waveform animate the selected source so you can verify before hitting Start. Mic-permission failures surface a one-click "Open mic settings" modal.

Live transcription — VAD-gated chunks through faster-whisper. A 1.5s speculative loop transcribes the audio tail so partial lines appear while you speak. UI re-adopts in-flight recordings across HMR / dev restarts. Live-tail scroll pinning: scrolled up reading? viewport stays put.

Speakers / Voices — pyannote 3.1 emits a 256-dim embedding per utterance. Tag-as-you-go (no upfront enrollment); the embedding folds into the Voice's pool, future chunks match via cosine distance. Unknowns are clustered live into Speaker 1/2/… rather than dumped into one bucket. Per-utterance re-assignment is fully lossless. Recompute re-runs diarization over a saved meeting against the current Voices DB. The Voices page lists every speaker with samples, snippets, merge / delete / rename / recolour, plus inline-editable email / org / role (surfaced in the People-note frontmatter, email also drives filename disambiguation when two voices share a display name).

Live AI coaching — debounced LLM call (~20s after last utterance, hard-cap 60s) extracts highlights + action items (yours and others'). A 2–5 bullet support-intelligence card suggests what to say next: tools to mention, counterarguments to pre-empt, clarifying questions. Live title refinement: same call returns an entity + topic so the meeting title sharpens as it progresses (freeze with the lock icon to stop overrides). A progress bar shows when the next refresh fires.

Obsidian vault — flat, generic layout. AuraScribe owns three folders; taxonomy beyond that is yours (add #customer/acme, folder conventions, MOCs — whatever fits your PKM):

<vault>/
  Meetings/YYYY/YYYY-MM-DD/HH-MM - <title>.md
  People/<Display Name>.md            # voice_id in frontmatter is the real identity key
  Daily/YYYY-MM/YYYY-MM-DD.md         # generated daily briefs, grouped by month

People notes are keyed by voice_id in frontmatter, not filename — rename a note in Obsidian and AuraScribe still finds it on the next write. Two voices with the same display name get a disambiguation suffix on creation (John Smith (acme).md — from email domain → org → short voice_id hash). Attendee wikilinks in meeting notes use alias syntax when disambiguated: [[John Smith (acme)|John Smith]].

Daily Brief — end-of-day rollup of every meeting on a date: tl;dr, highlights, decisions, open threads, action items, per-person takeaways, themes, tomorrow's focus. Marked stale and rebuilt in the background when a meeting finishes (toggleable).

Meeting Library — day-scoped browser: prev/next + native date picker match Daily Brief. Each card carries inline Summarize / Recompute / Delete icons; select multiple (or Select-all) and the bulk bar sweeps newest→oldest with an amber ring tracking the current card. Summarize is strictly sequential to avoid parallel LLM calls; Recompute fans out (DB-only). Live Feed keeps the last stopped meeting on screen until the next recording starts, so you never lose context mid-thought.

Meeting editing — rename (Obsidian file moves to match), trim before/after a timestamp, split at a timestamp (both halves get their own files), bulk-delete by id list or "last N days", click-to-play any utterance into the Opus recording at that exact offset, summarize + suggest title in one LLM pass.

Editable prompts — live_intelligence.md, daily_brief.md, meeting_analysis.md live under APP_DATA/prompts/ (seeded from bundled defaults). One-click open from Settings; edits picked up on the next call.

Diagnostics — React error boundary catches blank-screen renders; sidecar sys.excepthook writes crash-YYYYMMDD-HHMMSS.log; Rust panic hook writes crash-<unix>-rust.log; rotating file log (5 × 5 MB). Heartbeat polls /api/status every 30s — header flips to Sidecar unreachable if the process dies. On launch the Rust shell sweeps the process table and kills any orphan sidecars from a prior crash so the fresh child always owns the canonical port.

Everything sits under one data directory (default %APPDATA%\AuraScribe). The folder holds the SQLite DB, per-meeting .opus files, Whisper model cache, logs, editable prompts, and config.json. Copy it to migrate.

Advanced (Advanced Settings block — defaults match the old hard-coded values):

Settings UI shows the detected hardware (Detected: cuda · NVIDIA RTX 4090 · 24 GB VRAM) and tags each value custom / default. Most values need a sidecar restart; auto-capture knobs hot-reload. Sidecar bind: SIDECAR_HOST / SIDECAR_PORT env (default 127.0.0.1:8765).

DB policy — pre-GA. Schema changes drop and recreate (no migrations). db/database.py bumps a _CURRENT_SCHEMA_VERSION string; mismatch → fresh DB on next boot.

Tauri shell (Rust + WebView2 / WKWebView)
    │
    ├── React 19 UI (Vite + Tailwind + lucide-react)
    │       │
    │       └── HTTP + WebSocket → Python sidecar (127.0.0.1:8765)
    │                                    │
    │                                    ├── faster-whisper      (ASR; CUDA or CPU)
    │                                    ├── pyannote.audio 3.1  (diarization + 256-dim embeddings)
    │                                    ├── Silero VAD          (auto-capture + chunking)
    │                                    ├── sounddevice / soxr  (mic capture, Speex-AEC)
    │                                    ├── soundfile           (Opus recorder)
    │                                    ├── LLM client          (OpenAI-compat)
    │                                    ├── SQLite (aiosqlite)  (meetings, voices, embeddings)
    │                                    └── Obsidian writer     (flat Meetings/People/Daily markdown)
    │
    ├── spawns/kills sidecar subprocess
    └── panics → APPDATA\AuraScribe\logs\crash-*-rust.log

Sidecar binds to 127.0.0.1 only; CORS is locked to the Vite dev origin + Tauri webview schemes. On Windows, nvidia-*-cu12 bin/ directories are registered via os.add_dll_directory at import so CTranslate2 finds CUDA 12 runtimes without touching system PATH.

aurascribe/
├── .github/workflows/release.yml         CI: matrix CUDA/CPU builds → draft release
├── src-tauri/                            Rust shell (Tauri 2)
├── src/                                  React 19 + Vite + Tailwind
│   ├── App.tsx                           Root state, heartbeat poll, WS dispatch, page router
│   ├── components/                       Shell, Header, RecordingBar, TranscriptView,
│   │                                     VuMeter, Waveform, AutoCaptureChip, …
│   ├── pages/                            LiveFeed, MeetingLibrary, Review, Voices,
│   │                                     DailyBrief, Settings
│   └── lib/                              api.ts, useWebSocket, useLLMHealth, MicAudioContext
├── sidecar/                              Python 3.13 backend
│   ├── main.py                           uvicorn entry + crash excepthook
│   ├── build.ps1                         PyInstaller driver (auto-aligns torch wheel to host GPU)
│   └── aurascribe/
│       ├── api.py                        FastAPI app, CORS, lifespan, /ws, /api/status
│       ├── config.py                     Hardware probe, data-dir + config.json loader
│       ├── auto_capture.py               VAD-driven start/stop monitor + state machine
│       ├── meeting_manager.py            Recording lifecycle, speculative loop, finalize
│       ├── routes/                       meetings, voices, settings, daily_brief, intel,
│       │                                 plus _shared (singletons + helpers)
│       ├── audio/                        sounddevice capture, Silero VAD, Opus recorder
│       ├── transcription/                faster-whisper + pyannote, voice-pool matching
│       ├── llm/                          client, prompts, analysis, realtime (live intel
│       │                                 + title refinement), daily_brief,
│       │                                 plus user-editable .md prompts
│       ├── obsidian/writer.py            Meetings / People / Daily markdown writer
│       └── db/database.py                SQLite schema + drop-and-recreate version gate
├── package.json                          Scripts: dev, tauri:dev, tauri:build, build:sidecar, package
└── vite.config.ts                        Dev proxy + manual vendor chunks

Authoritative list lives in sidecar/aurascribe/routes/. Headlines:

Meetings (/api/meetings/*) — POST /start, POST /stop, GET / (paged), GET/DELETE/PATCH /{id}, /{id}/transcript, /{id}/audio (Opus, Range), /summarize, /suggest-title, /title-lock, /trim, /split, /rename-speaker, /utterances/{uid}/assign, /intel/refresh, /recompute, POST /bulk-delete, DELETE /all?days=.

Voices (/api/voices/*) — list, detail, rename / recolour / delete, snippet-delete, POST /merge.

Settings (/api/settings/*) — GET/PUT /data-dir, GET/PUT /config (response carries requires_restart).

Daily Brief (/api/daily-brief/*) — GET /?date=, POST /refresh?date=.

Intel (/api/intel/*) — GET /prompts, POST /open-prompt, GET /prompt-path.

System — GET /api/status, GET /api/models, GET/PUT /api/auto-capture, POST /api/system/open-mic-settings.

WebSocket /ws — push channel: utterances, partial_utterance, status, audio_level, realtime_intelligence, title_updated, auto_capture, daily_brief_updated.

npm run build:sidecar   # PyInstaller onedir bundle (10-15 min, ~1.5 GB)
npm run tauri:build     # Tauri build + NSIS installer
npm run package         # both, in sequence

build:sidecar auto-aligns your venv's torch wheel to the host GPU before PyInstaller runs (nvidia-smi detects GPU; mismatches trigger a pip install --index-url cu128 or cpu). CI pins the wheel upfront and skips this step.

CI — push a tag (git tag v0.2.0 && git push origin v0.2.0). .github/workflows/release.yml builds CUDA + CPU variants in parallel, splits the CUDA bundle into release-asset-sized parts, drafts a GitHub Release with both installers + the runtime zip + a picker table. You review and click Publish. workflow_dispatch runs builds without releasing.

Sidecar logs (incl. uvicorn) → stdout and %APPDATA%\AuraScribe\logs\sidecar.log (5 × 5 MB rotating). Crashes drop a sibling crash-YYYYMMDD-HHMMSS.log (Python) or crash-<unix>-rust.log (Rust). Useful grep prefixes: provisional:, voice-match:, diarize:, asr:, daily_brief:, auto-capture:, extras:. Frontend console: right-click → Inspect (devtools enabled in debug).

• Done — Tauri + React + sidecar; SQLite + Obsidian writer; live transcription + speculative partials; pyannote 3.1 diarization; tag-as-you-go Voices + provisional clustering + per-utterance re-assignment; recompute; live intel (highlights + action items + support coaching) with same-call title refinement + freeze; daily brief with auto-refresh; meeting trim / split; mid-recording re-adoption; NSIS installer + PyInstaller bundling; CI matrix for CUDA/CPU; hardware auto-detect + user-overridable; first-run welcome dialog; file-based logs + crash dumps; mic-permission detection; error boundaries; follow-tail scroll pinning; persistent globally-unique speaker colors + custom avatar upload; WASAPI loopback + Speex-AEC mix; macOS arm64 support; auto-capture (sustained-speech start, sustained-silence stop) with countdown-on-Stop-button + mid-recording toggle; generic Obsidian layout (Meetings/People/Daily) with voice_id-keyed People notes + filename disambiguation on display-name collisions; Meeting Library date picker + per-card and sweep-highlighted bulk actions; Live Feed continuity across stop→start; empty-meeting drop on stop; orphan sidecar cleanup at startup.
• Next — Code-signing for the installer (kill the SmartScreen "Unknown publisher" warning).
• Backlog — NVIDIA Parakeet-TDT 0.6B v3 as an optional ASR backend (NeMo on Windows is painful; deferred). Multi-vault routing. Scheduled meeting pre-briefs (calendar + prior-meeting context). Opt-in telemetry / crash uploader. MCP server (expose meetings / voices / intel to external agents). In-Settings mic test panel.

• Sidecar-only — .venv\Scripts\python sidecar\main.py; check http://127.0.0.1:8765/api/status.
• Frontend-only — npm run dev (Vite on 1420, proxies /api + /ws to 8765).
• Routes — api.py is just FastAPI wiring. Add a feature area: new routes/<name>.py with router = APIRouter(...), import in routes/__init__.py. Cross-router state lives in routes/_shared.py.
• Adding a config key — four edits: module-level getter in config.py, _CONFIG_KEYS in same file, _CONFIG_FIELDS + _effective_for + UserConfigUpdate in routes/settings.py, ConfigKey union + Settings UI field in the frontend.
• CUDA torch in dev — your venv defaults to the CPU torch wheel from PyPI. Run npm run build:sidecar once to auto-swap, or do it manually with the --index-url cu128 line above.
• Rust panics in dev — show as an error dialog + a crash file under %APPDATA%\AuraScribe\logs\. Release builds exit cleanly after the dialog.