diff --git a/linux_voice_assistant/__main__.py b/linux_voice_assistant/__main__.py
index 38526d4..1a9b069 100644
--- a/linux_voice_assistant/__main__.py
+++ b/linux_voice_assistant/__main__.py
@@ -18,7 +18,7 @@ from getmac import get_mac_address  # type: ignore
 from pymicro_wakeword import MicroWakeWord, MicroWakeWordFeatures
 from pyopen_wakeword import OpenWakeWord, OpenWakeWordFeatures
 
-from .models import Preferences, ServerState, WakeWordType
+from .models import Preferences, ServerState, WakeWordType, initial_stop_word_threshold
 from .mpv_player import MpvMediaPlayer
 from .peripheral_api import LVAEvent, PeripheralAPIServer
 from .satellite import VoiceSatelliteProtocol
@@ -379,6 +379,10 @@ async def main() -> None:
     initial_volume = max(0.0, min(1.0, float(initial_volume)))
     preferences.volume = initial_volume
 
+    # Load stop word sensitivity from preferences on startup, and ensure it's between 0.0 and 1.0
+    initial_threshold = initial_stop_word_threshold(preferences.stop_word_sensitivity)
+    preferences.stop_word_sensitivity = initial_threshold
+
     if args.enable_thinking_sound:
         preferences.thinking_sound = 1
 
@@ -441,6 +445,7 @@ async def main() -> None:
         output_only=args.output_only,
         download_dir=args.download_dir,
         volume=initial_volume,
+        stop_word_threshold=initial_threshold,
         mic_volume=preferences.mic_volume,
         mic_auto_gain=preferences.mic_auto_gain,
         mic_noise_suppression=preferences.mic_noise_suppression,
diff --git a/linux_voice_assistant/models.py b/linux_voice_assistant/models.py
index 85b1673..7586158 100644
--- a/linux_voice_assistant/models.py
+++ b/linux_voice_assistant/models.py
@@ -265,3 +265,14 @@ class ServerState:
         self.preferences.mic_volume = volume_int
         _LOGGER.info("Saving mic_volume %s to %s", volume_int, self.preferences_path)
         self.save_preferences()
+
+
+def initial_stop_word_threshold(saved_sensitivity: Optional[float]) -> float:
+    """
+    Resolve the stop word probability cutoff to start from, clamped to 0.0-1.0.
+    :param saved_sensitivity: Value persisted in preferences, or None if it has never been set.
+    """
+    if saved_sensitivity is None:
+        return ServerState.stop_word_threshold
+
+    return max(0.0, min(1.0, float(saved_sensitivity)))
diff --git a/tests/unit/test_models.py b/tests/unit/test_models.py
index ef10016..974250c 100644
--- a/tests/unit/test_models.py
+++ b/tests/unit/test_models.py
@@ -282,3 +282,30 @@ class TestPersistMicNoise:
         with patch.object(state, "save_preferences") as mock_save:
             state.persist_mic_noise(4.0)
             mock_save.assert_called_once()
+
+
+# ---------------------------------------------------------------------------
+# initial_stop_word_threshold()
+# ---------------------------------------------------------------------------
+
+
+class TestInitialStopWordThreshold:
+    def test_saved_sensitivity_is_restored(self):
+        from linux_voice_assistant.models import initial_stop_word_threshold
+
+        assert initial_stop_word_threshold(0.9) == pytest.approx(0.9)
+
+    def test_falls_back_to_server_state_default_when_never_set(self):
+        from linux_voice_assistant.models import ServerState, initial_stop_word_threshold
+
+        assert initial_stop_word_threshold(None) == pytest.approx(ServerState.stop_word_threshold)
+
+    def test_clamped_to_one_when_above(self):
+        from linux_voice_assistant.models import initial_stop_word_threshold
+
+        assert initial_stop_word_threshold(1.5) == pytest.approx(1.0)
+
+    def test_clamped_to_zero_when_below(self):
+        from linux_voice_assistant.models import initial_stop_word_threshold
+
+        assert initial_stop_word_threshold(-0.5) == pytest.approx(0.0)