fix(realtime): ignore a late speechStart for a finished Muse turn

A duplicate speechStart for a turn that already stopped used to make that
closed turn active again, so the next turnless PUSH_TO_TALK transcript was
routed to the finished item and dropped.

litellm/llms/meta/realtime/transformation.py

@@ -495,6 +495,8 @@ class MuseEventTransformer:
 
     def _speech_start(self, message: Mapping[str, JsonValue]) -> _TurnState:
         turn: Final = self._turn(_required_turn_id(message, "speechStart"))
+        if turn.stopped:
+            return turn
         turn.started = True
         self._active_turn_id = turn.item_id
         return turn

tests/test_litellm/llms/meta/realtime/test_meta_realtime_transformation.py

@@ -407,6 +407,24 @@ def test_finished_turn_ignores_late_duplicates():
     )
 
 
+def test_late_duplicate_speech_start_does_not_capture_the_next_turnless_transcript():
+    transformer = MuseEventTransformer()
+    transformer.configure(MuseSessionConfig(MUSE_MODEL, "PUSH_TO_TALK", 24_000, ()))
+    transformer.transform(json.loads(_event("speechStart", turnId="turn-1")))
+    transformer.transform(json.loads(_event("speechComplete", turnId="turn-1", transcript="first")))
+
+    assert transformer.transform(json.loads(_event("speechStart", turnId="turn-1"))) == ()
+    events = transformer.transform(json.loads(_event("transcript", transcript="second", final=True)))
+
+    assert [event["type"] for event in events] == [
+        "input_audio_buffer.speech_started",
+        "input_audio_buffer.speech_stopped",
+        "conversation.item.input_audio_transcription.completed",
+    ]
+    assert events[2]["transcript"] == "second"
+    assert events[2]["item_id"] != "turn-1"
+
+
 def test_turn_memory_is_bounded_by_turn_limit():
     transformer = MuseEventTransformer(turn_limit=2)