diff --git a/voice.py b/voice.py
index 15a4f39..14efb71 100644
--- a/voice.py
+++ b/voice.py
@@ -113,6 +113,7 @@ class VoiceSession:
         self.session = None
         self._task = None
         self._http_session = None
+        self._stt_session = None
         self._caller_key: bytes | None = None
         self._caller_identity: str | None = None
         self._caller_all_keys: dict = {}  # {index: bytes} — all caller keys by index
@@ -196,7 +197,7 @@ class VoiceSession:
         self._task = asyncio.create_task(self._run())
 
     async def stop(self):
-        for obj in [self.session, self.lk_room, self._http_session]:
+        for obj in [self.session, self.lk_room, self._http_session, self._stt_session]:
             if obj:
                 try:
                     if hasattr(obj, "aclose"):
@@ -344,9 +345,10 @@ class VoiceSession:
 
             # Voice pipeline — George (British male, multilingual DE/EN)
             self._http_session = aiohttp.ClientSession()
+            self._stt_session = aiohttp.ClientSession()  # separate session avoids WS/HTTP conflicts
             voice_id = os.environ.get("ELEVENLABS_VOICE_ID", DEFAULT_VOICE_ID)
             self.session = AgentSession(
-                stt=elevenlabs.STT(api_key=ELEVENLABS_KEY, http_session=self._http_session),
+                stt=elevenlabs.STT(api_key=ELEVENLABS_KEY, http_session=self._stt_session),
                 llm=lk_openai.LLM(base_url=LITELLM_URL, api_key=LITELLM_KEY, model=self.model),
                 tts=elevenlabs.TTS(voice_id=voice_id, model="eleven_multilingual_v2",
                                    api_key=ELEVENLABS_KEY, http_session=self._http_session),