(xai realtime): pass voice parameter

livekit-plugins/livekit-plugins-xai/livekit/plugins/xai/realtime/realtime_model.py

@@ -9,9 +9,14 @@
     ConversationItemAdded,
     ConversationItemDeletedEvent,
     ConversationItemInputAudioTranscriptionCompletedEvent,
+    RealtimeAudioConfig,
+    RealtimeAudioConfigInput,
+    RealtimeAudioConfigOutput,
     RealtimeConversationItemFunctionCall,
+    RealtimeSessionCreateRequest,
 )
 from openai.types.realtime.realtime_audio_input_turn_detection import ServerVad
+from openai.types.realtime.session_update_event import SessionUpdateEvent
 
 from livekit.agents import llm
 from livekit.agents.metrics import RealtimeModelMetrics
@@ -119,6 +124,42 @@ async def aclose(self) -> None:
             self.emit("metrics_collected", metrics)
         await super().aclose()
 
+    def _wrap_session_update(
+        self, event_id: str, session: RealtimeSessionCreateRequest
+    ) -> SessionUpdateEvent | dict[str, Any]:
+        # xAI expects `voice` and `turn_detection` as top-level session fields, whereas the
+        # OpenAI base nests them under audio.output / audio.input. Relocate them on the typed
+        # request (RealtimeSessionCreateRequest is extra="allow") before the base serializes,
+        # so this single seam covers every session.update the base emits — initial config,
+        # reconnect, and update_options — with no duplicated logic and no extra events.
+        # Gating on model_fields_set keeps an untouched field from leaking as a top-level null.
+        audio = session.audio
+        if isinstance(audio, RealtimeAudioConfig):
+            output = audio.output
+            if isinstance(output, RealtimeAudioConfigOutput) and "voice" in output.model_fields_set:
+                # voice/turn_detection are set as extra fields (the model is extra="allow")
+                session.voice = output.voice  # type: ignore[attr-defined]
+                output.model_fields_set.discard("voice")
+            audio_input = audio.input
+            if (
+                isinstance(audio_input, RealtimeAudioConfigInput)
+                and "turn_detection" in audio_input.model_fields_set
+            ):
+                session.turn_detection = audio_input.turn_detection  # type: ignore[attr-defined]
+                audio_input.model_fields_set.discard("turn_detection")
+
+            # if the relocation emptied both sub-blocks (a voice/turn_detection-only update),
+            # drop the now-hollow audio key instead of sending audio={"input":{},"output":{}}
+            out_set = isinstance(output, RealtimeAudioConfigOutput) and bool(
+                output.model_fields_set
+            )
+            in_set = isinstance(audio_input, RealtimeAudioConfigInput) and bool(
+                audio_input.model_fields_set
+            )
+            if not out_set and not in_set:
+                session.model_fields_set.discard("audio")
+        return super()._wrap_session_update(event_id=event_id, session=session)
+
     def _create_tools_update_event(self, tools: list[llm.Tool]) -> dict[str, Any]:
         event = super()._create_tools_update_event(tools)