.

core/session.py

@@ -1140,7 +1140,7 @@ class CallSession:
         actually_spoken_chunks: list[str] = []
         assistant_audio_started = [False]
         transcript = ""
-        filler_started = False
+        filler_started = asyncio.Event()
         self._set_active_unanswered_turn(epoch=epoch, utterance_audio=utterance_audio)
         try:
             LOGGER.info(
@@ -1276,6 +1276,7 @@ class CallSession:
                     epoch=epoch,
                     tool_name=None,
                     started_monotonic=llm_started_monotonic,
+                    filler_started=filler_started,
                     suppress=_is_terminal_farewell(transcript) or _is_farewell_transcript(transcript),
                 ),
                 name=f"{self.session_id}-filler-delay-{epoch}",
@@ -1289,16 +1290,16 @@ class CallSession:
                         epoch,
                         event.name,
                         event.tool_call_id,
-                        filler_started,
+                        filler_started.is_set(),
                     )
                     if not filler_trigger_task.done():
                         filler_trigger_task.cancel()
-                    if not filler_started:
-                        filler_started = True
+                    if not filler_started.is_set():
                         self._start_filler_audio(
                             epoch=epoch,
                             tool_name=event.name,
                             started_monotonic=llm_started_monotonic,
+                            filler_started=filler_started,
                         )
                     continue
 
@@ -1837,6 +1838,7 @@ class CallSession:
         epoch: int,
         tool_name: str | None,
         started_monotonic: float,
+        filler_started: asyncio.Event,
         suppress: bool = False,
     ) -> None:
         if suppress:
@@ -1846,16 +1848,18 @@ class CallSession:
                 epoch=epoch,
                 tool_name=tool_name,
                 started_monotonic=started_monotonic,
+                filler_started=filler_started,
             )
             return
         await asyncio.sleep(FILLER_AUDIO_DELAY_MS / 1000.0)
         self._ensure_generation(epoch)
-        if self._active_answer_audio_started:
+        if self._active_answer_audio_started or filler_started.is_set():
             return
         self._start_filler_audio(
             epoch=epoch,
             tool_name=tool_name,
             started_monotonic=started_monotonic,
+            filler_started=filler_started,
         )
 
     def _start_filler_audio(
@@ -1864,9 +1868,19 @@ class CallSession:
         epoch: int,
         tool_name: str | None,
         started_monotonic: float,
+        filler_started: asyncio.Event,
     ) -> None:
+        if filler_started.is_set():
+            LOGGER.info(
+                "realtime session %s filler audio skipped: already started for current turn epoch=%s tool=%s",
+                self.session_id,
+                epoch,
+                tool_name or "generic",
+            )
+            return
         if self._filler_task is not None and not self._filler_task.done():
             return
+        filler_started.set()
         LOGGER.info(
             "realtime session %s filler audio scheduled: epoch=%s tool=%s",
             self.session_id,
@@ -1878,6 +1892,7 @@ class CallSession:
                 epoch=epoch,
                 tool_name=tool_name,
                 started_monotonic=started_monotonic,
+                filler_started=filler_started,
             ),
             name=f"{self.session_id}-filler-{epoch}",
         )
@@ -1888,9 +1903,11 @@ class CallSession:
         epoch: int,
         tool_name: str | None,
         started_monotonic: float,
+        filler_started: asyncio.Event,
     ) -> None:
         if self._filler_audio is None:
             LOGGER.info("realtime session %s filler audio skipped: no library", self.session_id)
+            filler_started.clear()
             return
         filler_pcm = self._filler_audio.pick(tool_name)
         if not filler_pcm:
@@ -1899,6 +1916,7 @@ class CallSession:
                 self.session_id,
                 tool_name or "generic",
             )
+            filler_started.clear()
             return
         first_audio_seen = False
         sent_chunks = 0