Requirements

• 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.

Features

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.

Configuration

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.

Architecture

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.