bundles: meeting recorder with local transcription

bundles/meeting-recorder/README.md

@@ -0,0 +1,55 @@
+# Meeting Recorder
+
+Record a meeting in the browser and transcribe it on the same machine that served the page.
+
+The panel captures two audio sources, the meeting itself (a shared tab or window) and the
+microphone, mixes them in WebAudio, and uploads Opus every 15 seconds. On stop, a detached worker
+converts the audio, sends it to a local OpenAI-compatible transcription endpoint in ten-minute
+slices, and writes a timestamped markdown transcript. Nothing leaves the host, and no API key is
+involved.
+
+## What it needs
+
+- **ffmpeg and ffprobe** on the host.
+- **A transcription endpoint.** The `faster-whisper-server` bundle is the intended pairing: CPU,
+  int8, loopback `:8004`, which is this bundle's default. Any OpenAI-compatible
+  `/v1/audio/transcriptions` endpoint works.
+- **A secure context.** Browsers hand over tab audio only over HTTPS or on localhost. Reach the
+  dashboard through Tailscale Serve, a TLS reverse proxy, or `http://localhost`. The panel says so
+  on screen when the context is insecure, before you record silence by accident.
+
+## Configuration
+
+| Variable | Default | Meaning |
+|---|---|---|
+| `WHISPER_URL` | `http://localhost:8004/v1/audio/transcriptions` | the transcription endpoint |
+| `WHISPER_MODEL` | `Systran/faster-whisper-large-v3` | model name sent with each slice |
+| `WHISPER_SLICE_SECONDS` | `600` | slice length; smaller means finer progress, more requests |
+| `MEETING_RECORDER_EXPORT_DIR` | unset | if set, every transcript also lands in `<dir>/<date>-<slug>/transcript.md` |
+
+## Where recordings live
+
+`$CROW_HOME/data/meeting-recorder/<session-id>/`
+
+| File | What |
+|---|---|
+| `audio.webm` (or `audio.<ext>` for an upload) | the recording |
+| `audio.wav` | 16 kHz mono, what the transcriber read |
+| `meta.json` | title, timings, state, results |
+| `transcript.json` | segments with start, end, text |
+| `transcript.md` | the readable transcript, with any notes taken while listening |
+
+## Throughput
+
+Roughly 3.5x real time on an AMD Ryzen AI Max+ 395 with faster-whisper large-v3 int8 on CPU: a
+ninety-minute meeting finishes about twenty-five minutes after it ends. Transcription starts when
+recording stops; there is no live transcript.
+
+## Limits worth knowing before you rely on it
+
+- **No speaker labels.** Diarization is a second model and is not here. Every transcript carries a
+  line saying so, because a machine transcript with confident-looking text invites quotation.
+- **Names get misheard.** Verify any quotation against the audio before it travels.
+- **Recording other people carries obligations this bundle does not handle.** Many hosts prohibit
+  recording their sessions, and consent rules vary by jurisdiction. That judgement is the
+  operator's, before pressing record.

bundles/meeting-recorder/manifest.json

@@ -0,0 +1,45 @@
+{
+  "id": "meeting-recorder",
+  "name": "Meeting Recorder",
+  "version": "1.0.0",
+  "description": "Record a meeting in the browser (shared tab audio plus microphone) and transcribe it locally with faster-whisper. Nothing leaves the machine.",
+  "type": "bundle",
+  "author": "Crow",
+  "category": "ai",
+  "tags": ["audio", "transcription", "whisper", "meetings", "notes"],
+  "icon": "mic",
+  "panel": {
+    "id": "meeting-recorder",
+    "name": "Meeting Recorder",
+    "icon": "mic",
+    "route": "/dashboard/meeting-recorder",
+    "navOrder": 17
+  },
+  "panelRoutes": "panel/routes.js",
+  "requires": {
+    "env": [],
+    "min_ram_mb": 64,
+    "min_disk_mb": 500
+  },
+  "env_vars": [
+    {
+      "name": "WHISPER_URL",
+      "description": "OpenAI-compatible transcription endpoint. Defaults to the faster-whisper-server bundle on loopback :8004.",
+      "required": false,
+      "default": "http://localhost:8004/v1/audio/transcriptions"
+    },
+    {
+      "name": "WHISPER_MODEL",
+      "description": "Model name passed to that endpoint.",
+      "required": false,
+      "default": "Systran/faster-whisper-large-v3"
+    },
+    {
+      "name": "MEETING_RECORDER_EXPORT_DIR",
+      "description": "Optional. A directory that also receives a dated markdown copy of every transcript, for grepping or for a git repo.",
+      "required": false,
+      "default": ""
+    }
+  ],
+  "notes": "Needs ffmpeg on the host and a transcription endpoint (the faster-whisper-server bundle is the intended pairing; it is CPU-only and loopback-bound). No port of its own: the capture page and its upload endpoints ride the gateway through the panel and panelRoutes, so recordings inherit the dashboard's session auth. A browser only hands over tab audio in a secure context, so reach the dashboard over HTTPS (Tailscale Serve, a reverse proxy, or localhost)."
+}