bundles: meeting recorder with local transcription

[kh0pper]

Records a meeting in the browser and transcribes it on the same host. Built because a meeting worth referencing later should not have to pass through a cloud transcription service to become text.

What it does

The panel asks the browser for two audio sources, the meeting itself (a shared tab or window, which is the only way a page is allowed to hear another tab) and the microphone, mixes them in WebAudio, and uploads Opus every 15 seconds. A crashed tab or a closed laptop costs at most those 15 seconds. On stop, a detached worker converts the audio to 16 kHz mono, cuts it into ten-minute slices, sends each to a local OpenAI-compatible transcription endpoint, and writes a timestamped markdown transcript with any notes taken while listening.

It also takes a recording made some other way: a call app's own local recording, a phone voice memo, an old meeting. Same path from upload onward.

Shape

• No port of its own. The page and its upload endpoints ride the gateway through panel + panelRoutes, so recordings inherit the dashboard session rather than needing their own auth. No entry in the port registry.
• Bodies stream to disk. Audio arrives as application/octet-stream, which the gateway's global 1 MB JSON parser ignores, so a long upload never passes through memory.
• Auth is path-scoped (router.use("/dashboard/meeting-recorder-api", authMiddleware)), per the mount-time check on unpathed router middleware.
• Slice timestamps are offset back into meeting time before anything is written, so a two-hour transcript reads correctly end to end.

Requires

ffmpeg on the host, and a transcription endpoint. The faster-whisper-server bundle is the intended pairing and is the default (http://localhost:8004, CPU, int8, loopback). Browsers only hand over tab audio in a secure context, so the dashboard has to be reached over HTTPS or localhost; the panel says so on screen when it is not, rather than letting someone record silence.

Tested

Routes mounted on a bare Express app with a stub auth middleware, against a real faster-whisper server:

• three separately uploaded chunks reassembled and transcribed correctly (71 words, identical to the single-chunk result)
• an uploaded .m4a took the same path to the same output
• session listing and status polling

Throughput measured at about 3.5x real time with large-v3 int8 on CPU, so a ninety-minute meeting finishes roughly twenty-five minutes after it ends.

Deliberately not included

Speaker labels (diarization is a second model), a live transcript (transcription starts on stop), and any automatic summary. Every transcript carries a line saying it is a machine transcript whose names get misheard, because confident-looking text invites quotation.