✨(summary) make ASR response format configurable

Add WHISPERX_RESPONSE_FORMAT (default: diarized_json) so OpenAI-compatible
STT endpoints that do not support diarized_json, such as OVHcloud AI
Endpoints, can be used with verbose_json.

CHANGELOG.md

@@ -10,6 +10,7 @@ and this project adheres to
 
 ### Added
 
+- ✨(summary) make ASR response format configurable
 - ✨(frontend) add jittered exponential backoff to LiveKit reconnections
 - ✨(frontend) tell participants when a recording is interrupted
 

docs/features/transcription.md

@@ -81,6 +81,7 @@ sequenceDiagram
 | whisperx_api_key         | Secret    | —                                                                     | API key for accessing WhisperX.                                                                                                                                         |
 | whisperx_base_url        | String    | `"https://api.whisperx.com/v1"`                                       | Base URL for the WhisperX API.                                                                                                                                          |
 | whisperx_asr_model       | String    | `"whisper-1"`                                                         | ASR model used for transcription.                                                                                                                                       |
+| whisperx_response_format | String    | `"diarized_json"`                                                     | Response format requested from the transcription API. Use `verbose_json` for endpoints without diarization support (e.g. OVHcloud AI Endpoints).                        |
 | webhook_max_retries      | Integer   | `2`                                                                   | Maximum retries for webhook requests.                                                                                                                                   |
 | webhook_status_forcelist | List[Int] | `[502, 503, 504]`                                                     | HTTP status codes triggering webhook retry.                                                                                                                             |
 | webhook_backoff_factor   | Float     | `0.1`                                                                 | Exponential backoff factor for webhook retries.                                                                                                                         |

src/summary/summary/core/celery_worker.py

@@ -173,8 +173,9 @@ def transcribe_audio(
             # This is in particular true for word_segments and words.
             # WhisperX response is slightly different from OpenAI STT endpoints
             # response.
-            # At the same time "diarized_json" should be the value
-            # provided to STT endpoints in our context.
+            # At the same time "diarized_json" is the default value
+            # provided to STT endpoints in our context. It can be overridden
+            # for endpoints that do not support it (e.g. "verbose_json").
             url = urljoin(base_url.rstrip("/") + "/", "audio/transcriptions")
             res = requests.post(
                 url,
@@ -182,7 +183,7 @@ def transcribe_audio(
                     "model": settings.whisperx_asr_model,
                     "language": language,
                     "timestamp_granularities": ["word", "segment"],
-                    "response_format": "diarized_json",
+                    "response_format": settings.whisperx_response_format,
                 },
                 files={"file": audio_file},
                 headers={"Authorization": f"Bearer {api_key}"},

src/summary/summary/core/config.py

@@ -93,6 +93,9 @@ class Settings(BaseSettings):
     whisperx_api_key: SecretStr
     whisperx_base_url: Url = "https://api.openai.com/v1"
     whisperx_asr_model: str = "whisper-1"
+    # Use "verbose_json" for OpenAI-compatible STT endpoints that do not support
+    # "diarized_json" (e.g. OVHcloud AI Endpoints).
+    whisperx_response_format: str = "diarized_json"
     # ISO 639-1 language code (e.g., "en", "fr", "es")
     whisperx_default_language: Optional[str] = None
     whisperx_allowed_languages: Set[str] = {"en", "fr", "de", "nl"}

src/summary/tests/unit/test_validation_errors.py

@@ -221,3 +221,36 @@ def test_summary_task_retries_with_the_cleaned_exception(monkeypatch):
     error = retry.call_args.kwargs["exc"]
     assert isinstance(error, CleanedValidationError)
     assert CANARY.encode() not in pickle.dumps(error)
+
+
+@pytest.mark.parametrize(
+    ("configured", "expected"),
+    [(None, "diarized_json"), ("verbose_json", "verbose_json")],
+)
+def test_transcription_requests_configured_response_format(
+    monkeypatch, configured, expected
+):
+    """The ASR response format defaults to diarized_json and can be overridden."""
+    storage = MagicMock()
+    storage.prepare_audio_file.return_value.__enter__.return_value = (
+        Mock(),
+        {"duration": 1},
+    )
+    monkeypatch.setattr(celery_worker, "file_service", storage)
+    monkeypatch.setattr(celery_worker, "metadata_manager", Mock())
+    if configured is not None:
+        monkeypatch.setattr(
+            celery_worker,
+            "settings",
+            celery_worker.settings.model_copy(
+                update={"whisperx_response_format": configured}
+            ),
+        )
+    response = Mock(status_code=200)
+    response.json.return_value = {"text": "hello", "segments": []}
+    post = Mock(return_value=response)
+    monkeypatch.setattr(celery_worker.requests, "post", post)
+    celery_worker.transcribe_audio(
+        task_id="job", language="en", cloud_storage_url="https://example.com/audio"
+    )
+    assert post.call_args.kwargs["data"]["response_format"] == expected