fix(realtime): bill retained output and route Azure aliases

litellm/litellm_core_utils/realtime_streaming.py

@@ -450,7 +450,17 @@ class RealTimeStreaming:
                 output_seconds if isinstance(output_seconds, (int, float)) else synthetic_output_seconds
             )
             if isinstance(input_seconds, (int, float)) or resolved_output_seconds is not None:
-                if not self._should_store_message(event_obj):
+                if self._should_store_message(event_obj):
+                    if not isinstance(output_seconds, (int, float)) and synthetic_output_seconds is not None:
+                        self.messages.append(
+                            OpenAIRealtimeTranslationClosedEvent(
+                                type="session.closed",
+                                usage=OpenAIRealtimeTranslationDurationUsage(
+                                    type="duration", output_seconds=synthetic_output_seconds
+                                ),
+                            )
+                        )
+                else:
                     normalized_usage: Final = (
                         OpenAIRealtimeTranslationDurationUsage(
                             type="duration",

litellm/llms/azure/audio_transcriptions.py

@@ -43,18 +43,8 @@ class AzureAudioTranscription(AzureChatCompletion):
     ) -> TranscriptionResponse | Coroutine[Any, Any, TranscriptionResponse]:
         data: Final = {"model": model, "file": audio_file, **optional_params}
         sdk_data: Final = sdk_compatible_transcription_request_data(data)
-        model_info: Final = (
-            litellm.get_model_info(model=model, custom_llm_provider="azure")
-            if f"azure/{model}" in litellm.model_cost
-            else None
-        )
-        provider_specific_entry: Final = model_info.get("provider_specific_entry") if model_info is not None else None
         resolved_api_version: Final = (
-            litellm.AZURE_DEFAULT_API_VERSION
-            if provider_specific_entry is not None
-            and provider_specific_entry.get("transcription_deployment_api") == 1
-            and api_version in ("v1", "latest", "preview")
-            else api_version
+            litellm.AZURE_DEFAULT_API_VERSION if api_version in ("v1", "latest", "preview") else api_version
         )
 
         if atranscription is True:

litellm/model_prices_and_context_window_backup.json

@@ -234,7 +234,6 @@
         "mode": "audio_transcription",
         "provider_specific_entry": {
             "realtime_ga_only": 1,
-            "transcription_deployment_api": 1,
             "transcription_json_only": 1
         },
         "source": "https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-gpt-transcribe-and-gpt-live-transcribe-in-microsoft-foundry/4541740",

model_prices_and_context_window.json

@@ -234,7 +234,6 @@
         "mode": "audio_transcription",
         "provider_specific_entry": {
             "realtime_ga_only": 1,
-            "transcription_deployment_api": 1,
             "transcription_json_only": 1
         },
         "source": "https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-gpt-transcribe-and-gpt-live-transcribe-in-microsoft-foundry/4541740",

tests/test_litellm/litellm_core_utils/test_realtime_streaming.py

@@ -3027,6 +3027,28 @@ def test_translation_preserves_input_only_provider_usage(
     assert closed_events[0]["usage"] == expected_usage
 
 
+def test_translation_retained_input_only_close_event_bills_captured_output(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr(litellm, "logged_real_time_event_types", "*")
+    streaming = RealTimeStreaming(
+        websocket=MagicMock(),
+        backend_ws=MagicMock(),
+        logging_obj=MagicMock(),
+        model="gpt-realtime-translate",
+        translation_session=True,
+    )
+    streaming._translation_output_audio_bytes = 48000
+    close_event: Final = {"type": "session.closed", "usage": {"type": "duration", "input_seconds": 0.25}}
+
+    streaming._capture_translation_output_audio(close_event)
+    streaming.store_message(close_event)
+    streaming._finalize_translation_usage()
+
+    usage_events: Final = tuple(event["usage"] for event in streaming.messages if event.get("type") == "session.closed")
+    assert len(usage_events) == 2
+    assert sum(usage.get("input_seconds", 0.0) for usage in usage_events) == 0.25
+    assert sum(usage.get("output_seconds", 0.0) for usage in usage_events) == 1.0
+
+
 @pytest.mark.asyncio
 async def test_audio_delta_frame_parsed_at_most_once():
     client_ws = _beta_client_ws()

tests/unit/llms/openai/transcriptions/test_gpt_transcribe.py

@@ -261,14 +261,15 @@ def test_gpt_live_transcribe_rejects_file_transcription(local_model_cost_map: No
         ("2025-04-01-preview", "2025-04-01-preview"),
     ],
 )
-def test_azure_gpt_transcribe_resolves_api_version_in_provider(
-    local_model_cost_map: None, api_version: str | None, expected_api_version: str | None
+@pytest.mark.parametrize("model", ["gpt-transcribe", "custom-transcribe-deployment"])
+def test_azure_audio_transcription_resolves_api_version_in_provider(
+    model: str, api_version: str | None, expected_api_version: str | None
 ) -> None:
     handler = AzureAudioTranscription()
     handler.async_audio_transcriptions = MagicMock(return_value=MagicMock())
 
     handler.audio_transcriptions(
-        model="gpt-transcribe",
+        model=model,
         audio_file=io.BytesIO(b"audio"),
         optional_params={"stream": True},
         logging_obj=MagicMock(),