From 409de2fe2f0c4f62a8c4d9c907699deac00a069b Mon Sep 17 00:00:00 2001
From: SWAPI03 <swapnil300303@gmail.com>
Date: Fri, 4 Sep 2026 09:41:05 +0530
Subject: [PATCH] fix: label .webm as audio/webm in audio transcription

process_audio_file only ever handles audio for transcription, but it derived
the MIME type from the file extension via get_file_mime_type_from_extension,
which returns the video-biased "video/webm". Vertex Gemini transcription then
sends inlineData.mimeType: video/webm and returns an empty transcript for
audio-only webm uploads (browser MediaRecorder output).

Override the webm extension to audio/webm in this audio-only helper. Other
audio containers (ogg, m4a) already resolve to audio/*.

Fixes #38963
---
 .../litellm_core_utils/audio_utils/utils.py   | 13 +++++++++++-
 .../audio_utils/test_utils.py                 | 20 +++++++++++++++++++
 2 files changed, 32 insertions(+), 1 deletion(-)
 create mode 100644 tests/test_litellm/litellm_core_utils/audio_utils/test_utils.py

diff --git a/litellm/litellm_core_utils/audio_utils/utils.py b/litellm/litellm_core_utils/audio_utils/utils.py
index dab3e48f91a..ee1903db7fc 100644
--- a/litellm/litellm_core_utils/audio_utils/utils.py
+++ b/litellm/litellm_core_utils/audio_utils/utils.py
@@ -4,7 +4,9 @@ Utils used for litellm.transcription() and litellm.atranscription()
 
 import hashlib
 import os
+from collections.abc import Mapping
 from dataclasses import dataclass
+from types import MappingProxyType
 from typing import Final
 
 from litellm.types.files import (
@@ -16,6 +18,13 @@ from litellm.types.files import (
 )
 from litellm.types.utils import FileTypes
 
+# webm is an audio/video container; get_file_mime_type_from_extension returns
+# "video/webm", but this helper only ever handles audio for transcription, so a
+# .webm upload is audio. Labeling it video/webm makes Vertex Gemini
+# transcription return an empty transcript.
+# https://github.com/BerriAI/litellm/issues/38963
+_AUDIO_CONTAINER_MIME_OVERRIDES: Final[Mapping[str, str]] = MappingProxyType({"webm": "audio/webm"})
+
 
 @dataclass
 class ProcessedAudioFile:
@@ -121,7 +130,9 @@ def process_audio_file(audio_file: FileTypes) -> ProcessedAudioFile:
         try:
             # Extract extension from filename
             extension: Final = filename.split(".")[-1].lower() if "." in filename else "wav"
-            content_type = get_file_mime_type_from_extension(extension)
+            content_type = _AUDIO_CONTAINER_MIME_OVERRIDES.get(extension) or get_file_mime_type_from_extension(
+                extension
+            )
         except ValueError:
             # If extension is not recognized, fallback to audio/wav
             content_type = "audio/wav"
diff --git a/tests/test_litellm/litellm_core_utils/audio_utils/test_utils.py b/tests/test_litellm/litellm_core_utils/audio_utils/test_utils.py
new file mode 100644
index 00000000000..5da85d83d4b
--- /dev/null
+++ b/tests/test_litellm/litellm_core_utils/audio_utils/test_utils.py
@@ -0,0 +1,20 @@
+from litellm.litellm_core_utils.audio_utils.utils import process_audio_file
+
+
+def test_process_audio_file_labels_webm_as_audio():
+    # Regression for https://github.com/BerriAI/litellm/issues/38963: webm is an
+    # audio/video container and process_audio_file is audio-only, so a .webm
+    # upload must be audio/webm. video/webm makes Vertex Gemini transcription
+    # return an empty transcript
+    processed = process_audio_file(("speech.webm", b"\x1aE\xdf\xa3"))
+    assert processed.content_type == "audio/webm"
+
+
+def test_process_audio_file_keeps_known_audio_extension():
+    processed = process_audio_file(("speech.wav", b"RIFF"))
+    assert processed.content_type == "audio/wav"
+
+
+def test_process_audio_file_unknown_extension_falls_back_to_wav():
+    processed = process_audio_file(("recording.unknownext", b"x"))
+    assert processed.content_type == "audio/wav"