From 2cd0639fd66465650aee25b11148c5b4dad1f581 Mon Sep 17 00:00:00 2001
From: Emerson Gomes <emerson.gomes@thalesgroup.com>
Date: Tue, 22 Sep 2026 19:16:11 -0500
Subject: [PATCH] fix(vertex-ai): preserve nested speech locale and voice

---
 .../text_to_speech/transformation.py          | 12 +++++++++--
 .../text_to_speech/test_transformation.py     | 21 ++++++++++++++++---
 2 files changed, 28 insertions(+), 5 deletions(-)

diff --git a/litellm/llms/vertex_ai/text_to_speech/transformation.py b/litellm/llms/vertex_ai/text_to_speech/transformation.py
index 48598411670..2b799145422 100644
--- a/litellm/llms/vertex_ai/text_to_speech/transformation.py
+++ b/litellm/llms/vertex_ai/text_to_speech/transformation.py
@@ -245,10 +245,13 @@ class VertexAITextToSpeechConfig(BaseTextToSpeechConfig, VertexBase):
         self,
         voice: dict,  # mutable-ok: provider voice payload arrives as a concrete dictionary
     ) -> str | None:
-        voice_name: Final = self._get_str_value(voice, "name", "voiceName", "voice_name")
+        voice_name: Final = self._get_str_value(voice, "name", "voiceName", "voice_name", "voice")
         if voice_name is not None:
             return voice_name
         speech_config: Final = self._get_dict_value(voice, "speechConfig", "speech_config") or voice
+        nested_voice_name: Final = self._get_str_value(speech_config, "name", "voiceName", "voice_name", "voice")
+        if nested_voice_name is not None:
+            return nested_voice_name
         voice_config: Final = self._get_dict_value(speech_config, "voiceConfig", "voice_config")
         if voice_config is None:
             return None
@@ -273,7 +276,12 @@ class VertexAITextToSpeechConfig(BaseTextToSpeechConfig, VertexBase):
                 "name": voice,
             }
 
-        language_code: Final = self._get_str_value(voice, "languageCode", "language_code") or self.DEFAULT_LANGUAGE_CODE
+        speech_config: Final = self._get_dict_value(voice, "speechConfig", "speech_config") or voice
+        language_code: Final = (
+            self._get_str_value(voice, "languageCode", "language_code")
+            or self._get_str_value(speech_config, "languageCode", "language_code")
+            or self.DEFAULT_LANGUAGE_CODE
+        )
         model_name: Final = model
         speaker_configs: Final = self._extract_gemini_tts_speaker_configs(voice)
         if speaker_configs:
diff --git a/tests/unit/llms/vertex_ai/text_to_speech/test_transformation.py b/tests/unit/llms/vertex_ai/text_to_speech/test_transformation.py
index e85c81807bb..9c31a8a67d2 100644
--- a/tests/unit/llms/vertex_ai/text_to_speech/test_transformation.py
+++ b/tests/unit/llms/vertex_ai/text_to_speech/test_transformation.py
@@ -190,6 +190,23 @@ class TestVertexAITextToSpeechConfig:
         assert optional_params["audioEncoding"] == "LINEAR16"
         assert optional_params["vertex_voice_dict"] == expected_voice
 
+    def test_gemini_tts_nested_language_and_voice_alias_reach_cloud_tts(self):
+        config: Final = VertexAITextToSpeechConfig()
+        voice: Final = {"speech_config": {"language_code": "fr-FR"}, "voice": "Kore"}
+
+        voice_name, optional_params = config.map_openai_params(
+            model="gemini-3.1-flash-tts-preview",
+            optional_params={"response_format": "mp3"},
+            voice=voice,
+        )
+
+        assert voice_name is None
+        assert optional_params["vertex_voice_dict"] == {
+            "languageCode": "fr-FR",
+            "modelName": "gemini-3.1-flash-tts-preview",
+            "name": "Kore",
+        }
+
     @pytest.mark.parametrize(
         "voice",
         [
@@ -481,9 +498,7 @@ class TestVertexAILyriaTextToSpeechConfig:
         )
 
     def test_get_complete_url_encodes_injected_predict_path_segments(self, monkeypatch: pytest.MonkeyPatch) -> None:
-        injected: Final = (
-            "victim-project/locations/us-central1/publishers/google/models/other-model:predict?ignored="
-        )
+        injected: Final = "victim-project/locations/us-central1/publishers/google/models/other-model:predict?ignored="
         encoded: Final = (
             "victim-project%2Flocations%2Fus-central1%2Fpublishers%2Fgoogle"
             "%2Fmodels%2Fother-model%3Apredict%3Fignored%3D"