chore: snapshot livekit-cameras after serial pin + speechbrain lengths fix

prefetch_model.py

@@ -0,0 +1,49 @@
+"""One-time download of the speechbrain enhancement model.
+
+Each worker loads the model from the local cache, so run this once before
+starting all 20 workers:
+
+    .venv/bin/python prefetch_model.py
+"""
+
+import logging
+import sys
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
+log = logging.getLogger("prefetch")
+
+
+def main() -> int:
+    from config import load_config
+    cfg = load_config()
+    import torch
+    from speechbrain.inference.enhancement import SpectralMaskEnhancement
+
+    log.info("loading %s (first run downloads weights) ...", cfg.enhance_model)
+    model = SpectralMaskEnhancement.from_hparams(
+        source=cfg.enhance_model,
+        savedir="/root/.cache/speechbrain-enhancement",
+    )
+    model.mods.enhance_model.eval()
+
+    # quick self-test: 0.5 s of 16 kHz sine + noise
+    import numpy as np
+    n = int(cfg.audio_rate * 0.5)
+    t = np.arange(n) / cfg.audio_rate
+    x = (0.3 * np.sin(2 * np.pi * 440 * t) + 0.05 * np.random.randn(n))
+    x = x / np.abs(x).max()
+    with torch.no_grad():
+        w = torch.from_numpy(x.astype(np.float32)).unsqueeze(0).to(model.device)
+        # lengths are relative (1.0 = full sequence); older speechbrain
+        # builds don't accept the kwarg at all -> TypeError fallback.
+        try:
+            out = model.enhance_batch(w, lengths=torch.tensor([1.0]))
+        except TypeError:
+            out = model.enhance_batch(w)
+    log.info("self-test OK: in %d samples -> out %d samples",
+             n, out.numel())
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())