docs: serial identity, enhance hop/window, optional systemd unit

README.md

@@ -8,6 +8,9 @@ publishes:
 - a camera video track (1280x720 @ 30 fps by default, MJPG from V4L2)
 - a microphone audio track (16 kHz mono, cleaned with speechbrain)
 
+Identity is the USB serial, pinned by udev (`/dev/camNN` + `slots.json`).
+`--cameras 1,5` means participants `cam-01` and `cam-05`, not discovery order.
+
 ## Layout
 
 ```
@@ -20,6 +23,9 @@ config.py           .env / environment configuration
 prefetch_model.py   one-time speechbrain model download
 generate_udev_rules.py  (re)generate udev pinning + slots.json from live state
 slots.json          stable slot map: cam-NN -> USB serial, /dev/camNN, ALSA card
+smoke_test.py       one-camera LiveKit publish with enhancement ON
+scripts/bench_enhance.py  RTF bench (1 s window / 200 ms hop)
+deploy/livekit-cameras.service  optional systemd unit (do not enable blindly)
 ```
 
 ## Setup
@@ -52,6 +58,9 @@ sudo udevadm trigger --subsystem-match=video4linux --action=add
 sudo udevadm trigger --subsystem-match=sound --action=add
 ```
 
+The generator bakes the **current physical wiring** as canonical. Do not
+re-run `--write` unless you intend to remap slots.
+
 ## Run
 
 ```bash
@@ -62,11 +71,28 @@ sudo udevadm trigger --subsystem-match=sound --action=add
 
 ## Audio cleanup
 
-The mic audio is cleaned per chunk (default 200 ms) by
+The mic audio is cleaned per hop (default 200 ms) by
-`speechbrain/metricgan-plus-voicebank` (16 kHz, CPU). One worker process per
+`speechbrain/metricgan-plus-voicebank` (16 kHz, CPU) over a 1 s rolling
-camera loads the model once and reuses it; on this 28-core box all 20 stay
+window (`ENHANCE_CHUNK_S=1.0`, `ENHANCE_HOP_S=0.2`). One worker process
-real-time. `ENHANCE_MODE=off` (or a failed model load with `auto`) falls back
+per camera loads the model once and reuses it. `enhance_batch` must be
-to a lightweight high-pass + soft-clip DSP path so streams still work.
+called with `lengths=torch.tensor([1.0])` (relative, not sample count).
+`ENHANCE_MODE=off` (or a failed model load with `auto`) falls back to a
+lightweight high-pass + soft-clip DSP path so streams still work.
+
+Until the 1 s window is full, hops use the DSP fallback so LiveKit is
+not silent on startup.
+
+## systemd (optional, not enabled)
+
+```bash
+sudo cp deploy/livekit-cameras.service /etc/systemd/system/
+sudo systemctl daemon-reload
+# only if you want 24/7:
+sudo systemctl enable --now livekit-cameras
+```
+
+Requires a LiveKit server already listening (this host uses
+`/tmp/livekit-server --config /tmp/livekit-test.yaml` on :7880).
 
 ## Notes
 
@@ -74,6 +100,7 @@ to a lightweight high-pass + soft-clip DSP path so streams still work.
   -input_format mjpeg`) — cheap, and converts to BGRA for LiveKit.
 - Audio: ALSA mics on these C920s deliver 16 kHz **stereo** only; the worker
   downmixes to mono for LiveKit.
-- Camera <-> card pairing is resolved by matching USB port paths under /sys,
+- Camera identity is USB serial via `slots.json` / udev; USB port is
-  so each mic belongs to its own camera.
+  fallback only for cameras not yet in the map.
 - Workers auto-restart (up to 10 times) if a camera drops.
+- Local smoke: `.venv/bin/python smoke_test.py` (needs livekit-server on 7880).

deploy/livekit-cameras.service

@@ -0,0 +1,16 @@
+[Unit]
+Description=Publish C920 cameras (speechbrain audio) to LiveKit
+After=network-online.target sound.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+WorkingDirectory=/home/fdenkena/livekit-cameras
+ExecStart=/home/fdenkena/livekit-cameras/.venv/bin/python /home/fdenkena/livekit-cameras/run.py
+Restart=on-failure
+RestartSec=5
+Environment=PYTHONUNBUFFERED=1
+Environment=LIVEKIT_URL=ws://127.0.0.1:7880
+
+[Install]
+WantedBy=multi-user.target