diff --git a/app/core/orchestrator.py b/app/core/orchestrator.py
index cff898a..af2e6ac 100644
--- a/app/core/orchestrator.py
+++ b/app/core/orchestrator.py
@@ -1,3 +1,5 @@
+import asyncio
+
 from app.schemas import AudioChunk, PipelineTrace
 from app.pipeline.sentence_chunker import SentenceChunker
 from app.metrics import timer, metrics
@@ -197,49 +199,92 @@ class Orchestrator:
 
         parts: list[str] = []
         stream_fn = getattr(self.llm, "stream", None)
-        if stream_fn is not None:
-            async for delta in stream_fn(trace.cleaned_transcript or "", history=history, language=effective_language):
-                parts.append(delta)
-                if on_token:
-                    await on_token(delta)
-                if chunker:
-                    for sentence in chunker.feed(delta):
-                        await _emit_sentence(sentence)
-        else:
-            result = await self.llm.complete(trace.cleaned_transcript or "", history=history, language=effective_language)
-            parts.append(result)
-            if on_token:
-                await on_token(result)
-            if chunker:
-                for sentence in chunker.feed(result):
+
+        # Producer/Consumer: Die satzweise TTS-Synthese läuft in einem eigenen
+        # Consumer-Task, damit das Token-Streaming (Text in der Bubble) nicht
+        # mehr satzweise auf die Synthese wartet. Ein *einzelner* Consumer
+        # garantiert die Reihenfolge der Audio-Chunks — harte Invariante fürs
+        # gapless Playback im Client; out-of-order würde das Audio zerstören.
+        queue: asyncio.Queue | None = asyncio.Queue() if chunker else None
+        consumer_task: asyncio.Task | None = None
+
+        async def _consume() -> None:
+            while True:
+                sentence = await queue.get()
+                try:
+                    if sentence is None:  # Sentinel: keine Sätze mehr
+                        return
                     await _emit_sentence(sentence)
+                finally:
+                    queue.task_done()
 
-        trace.semantic_response = "".join(parts)
-        if not trace.semantic_response:
-            raise RuntimeError("LLM returned an empty response")
+        async def _dispatch(sentence: str) -> None:
+            # Satz in die Queue legen und sofort weiterlesen (kein await auf die
+            # Synthese). Ist der Consumer zuvor an einer Synthese-Exception
+            # gestorben, diese sofort hochreichen, statt weiter zu puffern.
+            if consumer_task is not None and consumer_task.done():
+                await consumer_task
+            await queue.put(sentence)
 
-        trace.spoken_response = await self.spoken_adapter.run(
-            trace.semantic_response,
-            language=effective_language,
-        )
-        trace.tts_ready_text = await self.tts_normalizer.run(
-            trace.spoken_response,
-            language=effective_language,
-            level=self.normalize_level,
-        )
+        if queue is not None:
+            consumer_task = asyncio.create_task(_consume())
 
-        if text_only:
-            return trace, b""
+        try:
+            if stream_fn is not None:
+                async for delta in stream_fn(trace.cleaned_transcript or "", history=history, language=effective_language):
+                    parts.append(delta)
+                    if on_token:
+                        await on_token(delta)
+                    if chunker:
+                        for sentence in chunker.feed(delta):
+                            await _dispatch(sentence)
+            else:
+                result = await self.llm.complete(trace.cleaned_transcript or "", history=history, language=effective_language)
+                parts.append(result)
+                if on_token:
+                    await on_token(result)
+                if chunker:
+                    for sentence in chunker.feed(result):
+                        await _dispatch(sentence)
 
-        if chunker:
-            tail = chunker.flush()
-            if tail:
-                await _emit_sentence(tail)
-            audio = b"".join(audio_parts)
-        else:
-            audio = await self.tts.synthesize(
-                trace.tts_ready_text, voice=voice, language=effective_language
+            trace.semantic_response = "".join(parts)
+            if not trace.semantic_response:
+                raise RuntimeError("LLM returned an empty response")
+
+            trace.spoken_response = await self.spoken_adapter.run(
+                trace.semantic_response,
+                language=effective_language,
             )
+            trace.tts_ready_text = await self.tts_normalizer.run(
+                trace.spoken_response,
+                language=effective_language,
+                level=self.normalize_level,
+            )
+
+            if text_only:
+                return trace, b""
+
+            if chunker:
+                tail = chunker.flush()
+                if tail:
+                    await _dispatch(tail)
+                await queue.put(None)   # Sentinel: Consumer beenden
+                await consumer_task     # auf restliche Synthese warten + Exceptions propagieren
+                consumer_task = None
+                audio = b"".join(audio_parts)
+            else:
+                audio = await self.tts.synthesize(
+                    trace.tts_ready_text, voice=voice, language=effective_language
+                )
+        finally:
+            # Bei Fehler/Abbruch den noch laufenden Consumer-Task abräumen,
+            # damit kein verwaister Task zurückbleibt.
+            if consumer_task is not None and not consumer_task.done():
+                consumer_task.cancel()
+                try:
+                    await consumer_task
+                except (asyncio.CancelledError, Exception):
+                    pass
 
         await self._emit_to_output(audio, output)
         return trace, audio