diff --git a/app/api/ws.py b/app/api/ws.py
index f11d1a6..17f6282 100644
--- a/app/api/ws.py
+++ b/app/api/ws.py
@@ -235,7 +235,7 @@ async def ws_voice(websocket: WebSocket, session_id: str | None = None, token: s
     fmt = "wav"
     active = None
     vad = None
-    vad_options: dict = {}
+    start_options: dict = {}  # Konfig aus dem start-Frame (Provider, audio_stream, language)
 
     async def _start_voice(audio: bytes, options: dict):
         nonlocal active
@@ -259,7 +259,7 @@ async def ws_voice(websocket: WebSocket, session_id: str | None = None, token: s
                     audio = bytes(audio_buffer)
                     audio_buffer.clear()
                     vad.reset()
-                    await _start_voice(audio, vad_options)
+                    await _start_voice(audio, start_options)
                 continue
 
             raw = message.get("text")
@@ -278,6 +278,7 @@ async def ws_voice(websocket: WebSocket, session_id: str | None = None, token: s
                 continue
             if ctype == "start":
                 audio_buffer.clear()
+                start_options = control  # Provider/audio_stream/language fuer den Turn merken
                 fmt = control.get("format", "wav")
                 if control.get("vad"):
                     vad = EnergyVAD(
@@ -285,7 +286,6 @@ async def ws_voice(websocket: WebSocket, session_id: str | None = None, token: s
                         threshold=control.get("vad_threshold", 500.0),
                         silence_ms=control.get("vad_silence_ms", 700.0),
                     )
-                    vad_options = control
                 else:
                     vad = None
                 continue
@@ -300,7 +300,8 @@ async def ws_voice(websocket: WebSocket, session_id: str | None = None, token: s
             audio_buffer.clear()
             if vad is not None:
                 vad.reset()
-            await _start_voice(audio, control)
+            # start-Frame-Konfig + end-Frame zusammenfuehren (end kann ueberschreiben)
+            await _start_voice(audio, {**start_options, **control})
     except WebSocketDisconnect:
         if active and not active.done():
             active.cancel()
diff --git a/tests/test_ws.py b/tests/test_ws.py
index 5619aa7..735bdfa 100644
--- a/tests/test_ws.py
+++ b/tests/test_ws.py
@@ -108,6 +108,22 @@ def test_ws_voice_transcribes_and_answers(monkeypatch):
         assert ws.receive_json()["type"] == "done"
 
 
+def test_ws_voice_start_frame_options_are_honored(monkeypatch):
+    # Provider stehen im START-Frame, der end-Frame ist leer -> muessen trotzdem gelten.
+    opts = _install_voice_stubs(monkeypatch)
+    with client.websocket_connect("/ws/voice") as ws:
+        ws.send_json({"type": "start", "format": "wav", **opts})
+        ws.send_bytes(b"PCMDATA")  # 7 Bytes
+        ws.send_json({"type": "end"})
+        transcript = ws.receive_json()
+        # Stub-STT (aus dem START-Frame) liefert "erkannt(<bytes>)" -> beweist: Override greift
+        assert transcript["type"] == "transcript" and transcript["text"] == "erkannt(7)"
+        assert ws.receive_json()["type"] == "ack"
+        assert ws.receive_json()["type"] == "semantic"
+        ws.receive_bytes()
+        assert ws.receive_json()["type"] == "done"
+
+
 def test_ws_voice_empty_buffer_errors(monkeypatch):
     opts = _install_voice_stubs(monkeypatch)
     with client.websocket_connect("/ws/voice") as ws: