From 21edacd9a525c4fef6938e1cb2bf45a1ab9c480b Mon Sep 17 00:00:00 2001
From: RealBhupesh <realbhupesh@gmail.com>
Date: Tue, 1 Sep 2026 11:34:20 +0530
Subject: [PATCH] fix(vertex): send audio/webm for Gemini transcription of webm

---
 .../gemini_transcribe_transformation.py       | 14 ++++++++++-
 ...tex_ai_gemini_transcribe_transformation.py | 25 +++++++++++++++++++
 2 files changed, 38 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/vertex_ai/audio_transcription/gemini_transcribe_transformation.py b/litellm/llms/vertex_ai/audio_transcription/gemini_transcribe_transformation.py
index f4db5eb110c..a923830c780 100644
--- a/litellm/llms/vertex_ai/audio_transcription/gemini_transcribe_transformation.py
+++ b/litellm/llms/vertex_ai/audio_transcription/gemini_transcribe_transformation.py
@@ -44,6 +44,7 @@ from litellm.types.utils import (
 
 DEFAULT_GEMINI_TRANSCRIBE_LOCATION: Final = "global"
 AUDIO_MODALITY: Final = "AUDIO"
+AMBIGUOUS_WEBM_MIME_TYPES: Final = frozenset({"application/webm", "application/x-webm", "video/webm", "video/x-webm"})
 
 
 class VertexGeminiAudioTranscriptionConfig(BaseAudioTranscriptionConfig, VertexBase):
@@ -151,6 +152,7 @@ class VertexGeminiAudioTranscriptionConfig(BaseAudioTranscriptionConfig, VertexB
         litellm_params: Mapping[str, object],
     ) -> AudioTranscriptionRequestData:
         processed_audio: Final = process_audio_file(audio_file)
+        mime_type: Final = _audio_transcription_mime_type(audio_file, processed_audio.content_type)
         request_body: Final = VertexGeminiTranscriptionRequest(
             contents=(
                 VertexGeminiTranscriptionContent(
@@ -158,7 +160,7 @@ class VertexGeminiAudioTranscriptionConfig(BaseAudioTranscriptionConfig, VertexB
                     parts=(
                         VertexGeminiTranscriptionPart(
                             inlineData=VertexGeminiTranscriptionInlineData(
-                                mimeType=processed_audio.content_type,
+                                mimeType=mime_type,
                                 data=base64.b64encode(processed_audio.file_content).decode("utf-8"),
                             )
                         ),
@@ -210,6 +212,16 @@ class VertexGeminiAudioTranscriptionConfig(BaseAudioTranscriptionConfig, VertexB
         return response
 
 
+def _audio_transcription_mime_type(audio_file: FileTypes, processed_content_type: str) -> str:
+    incoming_content_type: Final = audio_file[2] if isinstance(audio_file, tuple) and len(audio_file) >= 3 else None
+    if isinstance(incoming_content_type, str) and incoming_content_type.split(";", 1)[0].strip().lower().startswith(
+        "audio/"
+    ):
+        return incoming_content_type
+    processed_media_type: Final = processed_content_type.split(";", 1)[0].strip().lower()
+    return "audio/webm" if processed_media_type in AMBIGUOUS_WEBM_MIME_TYPES else processed_content_type
+
+
 def _audio_transcription_config(language: object) -> VertexGeminiTranscriptionAudioConfig:
     if not isinstance(language, str) or not language:
         return VertexGeminiTranscriptionAudioConfig()
diff --git a/tests/test_litellm/llms/vertex_ai/audio_transcription/test_vertex_ai_gemini_transcribe_transformation.py b/tests/test_litellm/llms/vertex_ai/audio_transcription/test_vertex_ai_gemini_transcribe_transformation.py
index eadd87d9c92..63a44457b02 100644
--- a/tests/test_litellm/llms/vertex_ai/audio_transcription/test_vertex_ai_gemini_transcribe_transformation.py
+++ b/tests/test_litellm/llms/vertex_ai/audio_transcription/test_vertex_ai_gemini_transcribe_transformation.py
@@ -1,6 +1,7 @@
 import base64
 import json
 import os
+from unittest.mock import Mock
 
 import httpx
 import pytest
@@ -217,6 +218,30 @@ class TestTransformRequest:
         assert round_tripped["generationConfig"] == {"audioTranscriptionConfig": {"languageCodes": ["en-US"]}}
         assert round_tripped["contents"][0]["role"] == "user"
 
+    def test_webm_audio_uses_audio_mime_type(self, config):
+        audio_file = Mock(spec=["name", "read", "seek"])
+        audio_file.name = "speech.webm"
+        audio_file.read.return_value = AUDIO_BYTES
+
+        request_data = config.transform_audio_transcription_request(
+            model="gemini-3.5-transcribe-preview",
+            audio_file=audio_file,
+            optional_params={},
+            litellm_params={},
+        )
+
+        assert request_data.data["contents"][0]["parts"][0]["inlineData"]["mimeType"] == "audio/webm"
+
+    def test_incoming_audio_mime_type_is_preserved(self, config):
+        request_data = config.transform_audio_transcription_request(
+            model="gemini-3.5-transcribe-preview",
+            audio_file=("speech.webm", AUDIO_BYTES, "audio/webm; codecs=opus"),
+            optional_params={},
+            litellm_params={},
+        )
+
+        assert request_data.data["contents"][0]["parts"][0]["inlineData"]["mimeType"] == "audio/webm; codecs=opus"
+
 
 class TestTransformResponse:
     def test_generate_content_response(self, config):