From 2fdb818327f070cfa96e175231df7ade46f01c1a Mon Sep 17 00:00:00 2001
From: Emerson Gomes <emerson.gomes@thalesgroup.com>
Date: Sat, 26 Sep 2026 08:52:05 -0500
Subject: [PATCH] fix(gemini): strip audio format from speech configuration

---
 .../vertex_and_google_ai_studio_gemini.py     |  4 +++-
 .../test_transformation.py                    | 20 +++++++++++++++++++
 2 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
index fa98f8f190d..ea0a14c3d1e 100644
--- a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
+++ b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
@@ -196,7 +196,9 @@ def normalize_gemini_speech_config(
         if isinstance(nested_config, Mapping)
         else value
     )
-    normalized: Final = _normalize_gemini_speech_config_item(config)
+    normalized: Final = _normalize_gemini_speech_config_item(
+        {key: item for key, item in config.items() if key != "format"}
+    )
     voice_name: Final = next(
         (normalized[key] for key in ("name", "voiceName", "voice") if isinstance(normalized.get(key), str)),
         None,
diff --git a/tests/unit/endpoints/speech/speech_to_completion_bridge/test_transformation.py b/tests/unit/endpoints/speech/speech_to_completion_bridge/test_transformation.py
index 6b01df20024..e65a82980e0 100644
--- a/tests/unit/endpoints/speech/speech_to_completion_bridge/test_transformation.py
+++ b/tests/unit/endpoints/speech/speech_to_completion_bridge/test_transformation.py
@@ -90,6 +90,26 @@ def test_gemini_request_keeps_voice_beside_nested_speech_config() -> None:
     }
 
 
+@pytest.mark.parametrize("wrapper", (None, "speech_config", "speechConfig"))
+def test_gemini_request_keeps_voice_and_language_without_audio_format(wrapper: str | None) -> None:
+    voice: Final = {"language_code": "fr-FR", "voice": "Kore", "format": "pcm16"}
+    request: Final = SpeechToCompletionBridgeTransformationHandler().transform_request(
+        model=GEMINI_TTS_MODEL,
+        input="Bonjour",
+        voice=voice if wrapper is None else {wrapper: voice},
+        optional_params={},
+        litellm_params={},
+        headers={},
+        litellm_logging_obj=MagicMock(),
+        custom_llm_provider="gemini",
+    )
+
+    assert request["audio"]["speech_config"] == {
+        "languageCode": "fr-FR",
+        "voiceConfig": {"prebuiltVoiceConfig": {"voiceName": "Kore"}},
+    }
+
+
 @pytest.mark.parametrize("response_format", ["mp3", "flac", "opus", "aac"])
 def test_gemini_tts_request_rejects_formats_gemini_cannot_produce(response_format: str) -> None:
     with pytest.raises(litellm.BadRequestError) as excinfo: