Fix bilingual initial greeting playback

core/session.py

@@ -1840,30 +1840,50 @@ class CallSession:
     ) -> None:
         first_audio_seen = False
         started_monotonic = time.perf_counter()
-        cached_greeting = (
-            self._filler_audio.pick("initial_greeting", allow_fallback=False)
-            if self._filler_audio is not None
-            else None
-        )
-        if cached_greeting:
-            await self._play_cached_initial_greeting(
-                epoch=epoch,
-                greeting_text=greeting_text,
-                pcm_audio=cached_greeting,
-                started_monotonic=started_monotonic,
-            )
-            return
         try:
             audio_chunk_count = 0
             audio_byte_count = 0
-            for segment_text, language_code in greeting_segments:
+            for segment_index, (segment_text, language_code) in enumerate(greeting_segments, start=1):
+                cached_segment = (
+                    self._filler_audio.pick(
+                        f"initial_greeting_{segment_index}_{language_code or 'default'}",
+                        allow_fallback=False,
+                    )
+                    if self._filler_audio is not None
+                    else None
+                )
+                if cached_segment:
+                    self._ensure_generation(epoch)
+                    audio_chunk_count += 1
+                    audio_byte_count += len(cached_segment)
+                    if not first_audio_seen:
+                        first_audio_seen = True
+                        self._log_latency(
+                            "initial_greeting_cached_ttfa",
+                            started_monotonic,
+                            epoch=epoch,
+                            message="session start -> cached initial greeting first audio",
+                        )
+                        self._set_state(SessionState.ASSISTANT_SPEAKING, reason="cached initial greeting started")
+                    LOGGER.info(
+                        "realtime session %s initial greeting cached segment: epoch=%s index=%s language=%s bytes=%s",
+                        self.session_id,
+                        epoch,
+                        segment_index,
+                        language_code,
+                        len(cached_segment),
+                    )
+                    await self.transport.send_audio(cached_segment)
+                    continue
+
                 async def one_shot_text_stream() -> AsyncGenerator[str, None]:
                     yield segment_text
 
                 LOGGER.info(
-                    "realtime session %s initial greeting segment start: epoch=%s language=%s chars=%s text=%r",
+                    "realtime session %s initial greeting segment start: epoch=%s index=%s language=%s chars=%s text=%r",
                     self.session_id,
                     epoch,
+                    segment_index,
                     language_code,
                     len(segment_text),
                     _preview_text(segment_text),