From adcdbc069595c0dda7f5e3bbdb11042f3b08f31c Mon Sep 17 00:00:00 2001
From: Florent Chehab <flo@nextmile.fr>
Date: Fri, 17 Jul 2026 17:35:30 +0200
Subject: [PATCH] =?UTF-8?q?=F0=9F=90=9B(summary)=20whisper=20call=20error?=
 =?UTF-8?q?=20handling?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* Consider that HTTP 400 errors are due to corrupted
audio files by  default.
* Properly reraise the http error otherwise so that
the retry mechanism actually works.
---
 CHANGELOG.md                              |  1 +
 src/summary/summary/core/celery_worker.py | 21 ++++++++++++++++++---
 src/summary/summary/core/file_service.py  |  6 ++++++
 src/summary/summary/core/shared_models.py |  8 +++++++-
 4 files changed, 32 insertions(+), 4 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 00538e55..d6da3440 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -34,6 +34,7 @@ and this project adheres to
 - 🩹(backend) identify externally provisioned users to PostHog
 - 🐛(backend) fix info panel crash for unregistered rooms
 - ♿️(frontend) focus side panel container on open #1452
+- 🐛(summary) whisper call error handling
 
 ## [1.23.0] - 2026-07-08
 
diff --git a/src/summary/summary/core/celery_worker.py b/src/summary/summary/core/celery_worker.py
index cbfa0bcf..11db588a 100644
--- a/src/summary/summary/core/celery_worker.py
+++ b/src/summary/summary/core/celery_worker.py
@@ -18,7 +18,12 @@ from requests import exceptions
 from summary.core.analytics import MetadataManager, get_analytics
 from summary.core.config import get_settings
 from summary.core.docs_service import create_document_in_lasuite_docs
-from summary.core.file_service import FileService, FileServiceException, TranscribeError
+from summary.core.file_service import (
+    CorruptedAudioFile,
+    FileService,
+    FileServiceException,
+    TranscribeError,
+)
 from summary.core.llm_service import LLMException, LLMObservability, LLMService
 from summary.core.locales import get_locale
 from summary.core.models import (
@@ -147,10 +152,20 @@ def transcribe_audio(
                 # Mimic OpenAI's timeout settings
                 timeout=(60, 10 * 60),
             )
+            if res.status_code == 400:
+                logger.info(
+                    "WhisperX transcription failed, "
+                    "likely due to a corrupted audio file: %s",
+                    res.text,
+                )
+                raise CorruptedAudioFile("WhisperX coudln't decode the audio file.")
+
             try:
                 res.raise_for_status()
-            except requests.exceptions.HTTPError as e:
-                raise RuntimeError("WhisperX transcription failed, %s", res.text) from e
+            except requests.exceptions.HTTPError:
+                logger.exception("WhisperX transcription failed")
+                # We reraise the error so that it can be retried by celery
+                raise
 
             transcription_json: dict[str, Any] = res.json()
             # We remove the "usage" key from the transcription_json dictionary
diff --git a/src/summary/summary/core/file_service.py b/src/summary/summary/core/file_service.py
index 920f577e..3c10f26d 100644
--- a/src/summary/summary/core/file_service.py
+++ b/src/summary/summary/core/file_service.py
@@ -41,6 +41,12 @@ class NoAudioInFileError(TranscribeError):
     error_code = "no_audio_in_file"
 
 
+class CorruptedAudioFile(TranscribeError):
+    """Raised when a media file does not contain any audio."""
+
+    error_code = "corrupted_audio_file"
+
+
 def _get_duration_from_packets(local_path: Path) -> float:
     """Estimate duration from audio packet timestamps."""
     # Run ffprobe to inspect the first audio stream in the file.
diff --git a/src/summary/summary/core/shared_models.py b/src/summary/summary/core/shared_models.py
index 2e80896c..f7423d7d 100644
--- a/src/summary/summary/core/shared_models.py
+++ b/src/summary/summary/core/shared_models.py
@@ -96,7 +96,13 @@ class TranscribeWebhookFailurePayload(BaseWebhook):
     # we authorized any other string than the one in the literal
     # to avoid causing a breaking change on the client side
     error_code: (
-        Literal["unknown_error", "no_audio_in_file", "media_duration_too_long"] | str
+        Literal[
+            "unknown_error",
+            "no_audio_in_file",
+            "media_duration_too_long",
+            "corrupted_audio_file",
+        ]
+        | str
     ) = Field(title="Error code", description="The error code.")