diff --git a/headmic.py b/headmic.py
index 4e54783..2888c26 100644
--- a/headmic.py
+++ b/headmic.py
@@ -469,8 +469,8 @@ async def startup():
     edgetpu_model_path = model_dir / "yamnet_edgetpu.tflite"
     model_path = model_dir / "yamnet.tflite"
     class_map_path = model_dir / "yamnet_class_map.csv"
-    # Edge TPU opt-in via env var (segfaults with some libedgetpu versions)
-    use_edgetpu = os.environ.get("USE_EDGETPU", "").lower() in ("1", "true", "yes") and edgetpu_model_path.exists()
+    # Auto-detect Edge TPU — probe in subprocess catches segfaults safely
+    use_edgetpu = edgetpu_model_path.exists()
     active_model = edgetpu_model_path if use_edgetpu else model_path
     if active_model.exists() and class_map_path.exists():
         try:
diff --git a/sound_id.py b/sound_id.py
index 4b9ea40..da631c0 100644
--- a/sound_id.py
+++ b/sound_id.py
@@ -85,6 +85,29 @@ CATEGORY_GROUPS = {
 
 
 class SoundClassifier:
+    @staticmethod
+    def _probe_edgetpu(model_path: str) -> bool:
+        """Test Edge TPU in a subprocess to catch segfaults safely."""
+        import subprocess, sys
+        try:
+            result = subprocess.run(
+                [sys.executable, "-c",
+                 "import ai_edge_litert.interpreter as tfl; "
+                 f"d = tfl.load_delegate('libedgetpu.so.1'); "
+                 f"i = tfl.Interpreter(model_path='{model_path}', experimental_delegates=[d]); "
+                 "i.allocate_tensors(); "
+                 "print('ok')"],
+                capture_output=True, text=True, timeout=10
+            )
+            if result.returncode == 0 and "ok" in result.stdout:
+                logger.info("Edge TPU probe: OK")
+                return True
+            logger.warning("Edge TPU probe failed: %s", result.stderr.strip() or f"exit {result.returncode}")
+            return False
+        except Exception as e:
+            logger.warning("Edge TPU probe error: %s", e)
+            return False
+
     def __init__(self, model_path, class_map_path, use_edgetpu=False):
         # Load class names
         self._class_names = []
@@ -105,14 +128,21 @@ class SoundClassifier:
         import ai_edge_litert.interpreter as tfl
 
         if use_edgetpu:
-            delegate = tfl.load_delegate("libedgetpu.so.1")
-            self._interp = tfl.Interpreter(
-                model_path=str(model_path),
-                experimental_delegates=[delegate],
-            )
-            logger.info("YAMNet loaded on Edge TPU")
-        else:
-            self._interp = tfl.Interpreter(model_path=str(model_path))
+            if self._probe_edgetpu(model_path):
+                delegate = tfl.load_delegate("libedgetpu.so.1")
+                self._interp = tfl.Interpreter(
+                    model_path=str(model_path),
+                    experimental_delegates=[delegate],
+                )
+                logger.info("YAMNet loaded on Edge TPU")
+            else:
+                logger.warning("Edge TPU probe failed (segfault or error) — falling back to CPU")
+                use_edgetpu = False
+
+        if not use_edgetpu:
+            # Use CPU model (swap edgetpu model path for CPU model if needed)
+            cpu_path = str(model_path).replace("_edgetpu.tflite", ".tflite")
+            self._interp = tfl.Interpreter(model_path=cpu_path)
             logger.info("YAMNet loaded on CPU")
 
         self._interp.allocate_tensors()