fix(gemini-realtime): keep the client's voice on Vertex AI native-audio Live

Vertex AI Live accepts a speechConfig on setup for native-audio models, so
stripping it silently discarded the voice a client asked for. Confirmed against
a live BidiGenerateContent session on gemini-live-2.5-flash-native-audio and on
gemini-live-2.5-flash-preview-native-audio-09-2025: both return setupComplete
with speechConfig present.

The strip stays in place for Google AI Studio, which was never verified to
accept it, via an overridable predicate matching the existing
_include_function_response_id pattern. The responseModalities TEXT to AUDIO
coercion is unchanged, since Vertex does reject TEXT on these models.

litellm/llms/gemini/realtime/transformation.py

@@ -92,6 +92,10 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
         """Google AI Studio Gemini 3.5+ accepts ``id`` on functionResponses; Vertex AI rejects it."""
         return True
 
+    def _strip_native_audio_speech_config(self) -> bool:
+        """Google AI Studio native-audio Live rejects ``speechConfig`` on setup; Vertex AI accepts it."""
+        return True
+
     @staticmethod
     def _usage_detail_alias(details: Any, defaults: dict[str, int]) -> dict[str, Any]:
         if not isinstance(details, dict):
@@ -382,8 +386,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
         without_text: Final = [modality for modality in normalized if modality != "TEXT"]
         return without_text if without_text else ["AUDIO"]
 
-    @staticmethod
-    def _finalize_gemini_live_setup(model: str, setup: dict[str, Any]) -> dict[str, Any]:
+    def _finalize_gemini_live_setup(self, model: str, setup: dict[str, Any]) -> dict[str, Any]:
         """Drop fields Gemini Live native-audio rejects on ``setup``."""
         generation_config: Final = setup.get("generationConfig")
         if isinstance(generation_config, dict):
@@ -392,7 +395,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
                 generation_config["responseModalities"] = GeminiRealtimeConfig._coerce_response_modalities(
                     model, modalities
                 )
-            if GeminiRealtimeConfig._is_native_audio_model(model):
+            if self._strip_native_audio_speech_config() and GeminiRealtimeConfig._is_native_audio_model(model):
                 generation_config.pop("speechConfig", None)
         return setup
 

litellm/llms/vertex_ai/realtime/transformation.py

@@ -35,6 +35,9 @@ class VertexAIRealtimeConfig(GeminiRealtimeConfig):
     def _include_function_response_id(self) -> bool:
         return False
 
+    def _strip_native_audio_speech_config(self) -> bool:
+        return False
+
     # ------------------------------------------------------------------
     # URL
     # ------------------------------------------------------------------

tests/test_litellm/llms/vertex_ai/realtime/test_vertex_ai_realtime_transformation.py

@@ -462,3 +462,53 @@ def test_vertex_function_call_output_omits_id():
     assert "id" not in function_response
     assert function_response["name"] == "terminate_call"
     assert function_response["response"] == {"status": "ok"}
+
+
+def test_vertex_native_audio_keeps_requested_voice(patch_native_audio_cost_map_entry):
+    """Regression: Vertex Live accepts speechConfig on native audio, so the client's voice must survive.
+
+    Stripping it silently dropped voice selection for every Vertex native-audio
+    session. TEXT is still coerced away, which Vertex does reject.
+    """
+    cfg = VertexAIRealtimeConfig(
+        access_token="tok", project="my-proj", location="us-central1"
+    )
+    session_update = {
+        "type": "session.update",
+        "session": {
+            "output_modalities": ["text"],
+            "audio": {"output": {"voice": "Aoede"}},
+        },
+    }
+
+    messages = cfg.transform_realtime_request(
+        json.dumps(session_update),
+        _NATIVE_AUDIO_MODEL,
+        session_configuration_request=None,
+    )
+
+    generation_config = json.loads(messages[0])["setup"]["generationConfig"]
+    assert generation_config["speechConfig"]["voiceConfig"]["prebuiltVoiceConfig"]["voiceName"] == "Aoede"
+    assert generation_config["responseModalities"] == ["AUDIO"]
+
+
+def test_google_ai_studio_native_audio_still_strips_voice(patch_native_audio_cost_map_entry):
+    """Google AI Studio's own native-audio Live is untouched: only Vertex was verified to accept speechConfig."""
+    from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig
+
+    messages = GeminiRealtimeConfig().transform_realtime_request(
+        json.dumps(
+            {
+                "type": "session.update",
+                "session": {
+                    "output_modalities": ["audio"],
+                    "audio": {"output": {"voice": "Aoede"}},
+                },
+            }
+        ),
+        _NATIVE_AUDIO_MODEL,
+        session_configuration_request=None,
+    )
+
+    generation_config = json.loads(messages[0])["setup"]["generationConfig"]
+    assert "speechConfig" not in generation_config