diff --git a/engine/core/duplex_pipeline.py b/engine/core/duplex_pipeline.py
index e360951..3fe5e26 100644
--- a/engine/core/duplex_pipeline.py
+++ b/engine/core/duplex_pipeline.py
@@ -72,6 +72,7 @@ class DuplexPipeline:
     _PCM_FRAME_BYTES = 640  # 16k mono pcm_s16le, 20ms
     _ASR_DELTA_THROTTLE_MS = 300
     _LLM_DELTA_THROTTLE_MS = 80
+    _ASR_CAPTURE_MAX_MS = 15000
     _DEFAULT_TOOL_SCHEMAS: Dict[str, Dict[str, Any]] = {
         "current_time": {
             "name": "current_time",
@@ -166,6 +167,7 @@ class DuplexPipeline:
         self._max_audio_buffer_bytes = int(settings.sample_rate * 2 * max_buffer_seconds)
         self._asr_start_min_speech_ms: int = settings.asr_start_min_speech_ms
         self._asr_capture_active: bool = False
+        self._asr_capture_started_ms: float = 0.0
         self._pending_speech_audio: bytes = b""
         # Keep a short rolling pre-speech window so VAD transition latency
         # does not clip the first phoneme/character sent to ASR.
@@ -1078,6 +1080,15 @@ class DuplexPipeline:
                 # 4. Check for End of Utterance - this triggers LLM response
                 if self.eou_detector.process(vad_status, force_eligible=self._asr_capture_active):
                     await self._on_end_of_utterance()
+                elif (
+                    self._asr_capture_active
+                    and self._asr_capture_started_ms > 0.0
+                    and (time.monotonic() * 1000.0 - self._asr_capture_started_ms) >= self._ASR_CAPTURE_MAX_MS
+                ):
+                    logger.warning(
+                        f"[EOU] Force finalize after ASR capture timeout: {self._ASR_CAPTURE_MAX_MS}ms"
+                    )
+                    await self._on_end_of_utterance()
                 elif (
                     vad_status == "Silence"
                     and not self.eou_detector.is_speaking
@@ -1180,6 +1191,7 @@ class DuplexPipeline:
             self._last_sent_transcript = ""
             self.eou_detector.reset()
             self._asr_capture_active = False
+            self._asr_capture_started_ms = 0.0
             self._pending_speech_audio = b""
 
             # Clear ASR buffer. Interim starts only after ASR capture is activated.
@@ -1211,6 +1223,7 @@ class DuplexPipeline:
             self._audio_buffer = capture_audio[-self._max_audio_buffer_bytes:]
 
         self._asr_capture_active = True
+        self._asr_capture_started_ms = time.monotonic() * 1000.0
         logger.debug(
             f"ASR capture started after speech gate ({self._asr_start_min_speech_ms}ms), "
             f"capture={len(capture_audio)} bytes"
@@ -1247,6 +1260,7 @@ class DuplexPipeline:
             self._audio_buffer = b""
             self._last_sent_transcript = ""
             self._asr_capture_active = False
+            self._asr_capture_started_ms = 0.0
             self._pending_speech_audio = b""
             # Return to idle; don't force LISTENING which causes buffering on silence
             await self.conversation.set_state(ConversationState.IDLE)
@@ -1272,6 +1286,7 @@ class DuplexPipeline:
         self._pending_transcript_delta = ""
         self._last_transcript_delta_emit_ms = 0.0
         self._asr_capture_active = False
+        self._asr_capture_started_ms = 0.0
         self._pending_speech_audio = b""
 
         # Process the turn - trigger LLM response
@@ -2017,6 +2032,7 @@ class DuplexPipeline:
         self._audio_buffer = b""
         self.eou_detector.reset()
         self._asr_capture_active = False
+        self._asr_capture_started_ms = 0.0
         self._pending_speech_audio = b""
 
     async def _stop_current_speech(self) -> None:
diff --git a/engine/processors/eou.py b/engine/processors/eou.py
index 22d104f..7a3758b 100644
--- a/engine/processors/eou.py
+++ b/engine/processors/eou.py
@@ -53,6 +53,18 @@ class EouDetector:
 
         if vad_status == "Silence":
             if not self.is_speaking:
+                # Fallback path: ASR has already started but VAD speaking state
+                # was lost (e.g. VAD flicker). Allow silence-only EOU so the
+                # turn can still close instead of stalling forever.
+                if not force_eligible:
+                    return False
+                if self.silence_start_time is None:
+                    self.silence_start_time = now
+                silence_duration = now - self.silence_start_time
+                if silence_duration >= self.threshold and not self.triggered:
+                    self.triggered = True
+                    self.silence_start_time = None
+                    return True
                 return False
             if self.silence_start_time is None:
                 self.silence_start_time = now