From 35a375e26f99f6246fd2e68e4c002532fef111a0 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 21:10:15 -0700
Subject: [PATCH] fix(speech): stop vertex gemini tts from dropping
 response_format in cloud tts param mapping

---
 litellm/utils.py                 |  4 ++++
 tests/test_litellm/test_utils.py | 20 ++++++++++++++++++++
 2 files changed, 24 insertions(+)

diff --git a/litellm/utils.py b/litellm/utils.py
index 5e9e115ed54..e7028301dab 100644
--- a/litellm/utils.py
+++ b/litellm/utils.py
@@ -9409,6 +9409,10 @@ class ProviderConfigManager:
 
             return RunwayMLTextToSpeechConfig()
         elif litellm.LlmProviders.VERTEX_AI == provider:
+            if "gemini" in model:
+                # Gemini TTS uses the speech_to_completion bridge, and Google Cloud TTS param
+                # mapping would drop response_format before the bridge sees it (LIT-6501)
+                return None
             from litellm.llms.vertex_ai.text_to_speech.transformation import (
                 VertexAITextToSpeechConfig,
             )
diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py
index 6524353aa48..935e4c61535 100644
--- a/tests/test_litellm/test_utils.py
+++ b/tests/test_litellm/test_utils.py
@@ -1416,6 +1416,26 @@ def test_get_provider_rerank_config():
     assert isinstance(config, HostedVLLMRerankConfig)
 
 
+def test_get_provider_text_to_speech_config_vertex_gemini_skips_cloud_tts():
+    """Regression for LIT-6501: mapping vertex Gemini TTS params through Google Cloud TTS
+    dropped response_format before the speech_to_completion bridge could honor it."""
+    from litellm.llms.vertex_ai.text_to_speech.transformation import VertexAITextToSpeechConfig
+    from litellm.utils import LlmProviders
+
+    assert (
+        ProviderConfigManager.get_provider_text_to_speech_config(
+            model="gemini-2.5-flash-preview-tts", provider=LlmProviders.VERTEX_AI
+        )
+        is None
+    )
+    assert isinstance(
+        ProviderConfigManager.get_provider_text_to_speech_config(
+            model="en-US-Studio-O", provider=LlmProviders.VERTEX_AI
+        ),
+        VertexAITextToSpeechConfig,
+    )
+
+
 # Models that should be skipped during testing
 OLD_PROVIDERS = ["aleph_alpha", "palm"]
 SKIP_MODELS = [