Fix: wrap sync generator in executor, not async for

main.py

@@ -166,7 +166,7 @@ def get_custom_voices() -> List[str]:
 
 async def generate_speech(text: str, voice: str) -> bytes:
     """
-    Generate speech using Orpheus model (async version).
+    Generate speech using Orpheus model (async wrapper).
     
     Args:
         text: Text to convert (may include emotion tags)
@@ -191,7 +191,8 @@ async def generate_speech(text: str, voice: str) -> bytes:
     
     print(f"{text}")
     
-    # Generate speech using Orpheus - async iteration!
+    # Run synchronous generation in thread pool to not block event loop
+    def _generate_sync():
         import numpy as np
         audio_chunks = []
         
@@ -200,8 +201,8 @@ async def generate_speech(text: str, voice: str) -> bytes:
             voice=voice,
         )
         
-    # Async iteration over the generator
+        # Sync iteration - generator yields audio chunks
-    async for audio_chunk in syn_tokens:
+        for audio_chunk in syn_tokens:
             audio_chunks.append(audio_chunk)
         
         # Combine chunks into single audio
@@ -217,6 +218,10 @@ async def generate_speech(text: str, voice: str) -> bytes:
         
         return buffer.getvalue()
     
+    # Run in executor to avoid blocking
+    loop = asyncio.get_event_loop()
+    return await loop.run_in_executor(None, _generate_sync)
+
 
 def save_audio_to_file(job_id: str, audio_bytes: bytes) -> str:
     """Save audio bytes to WAV file."""
@@ -528,15 +533,15 @@ async def stream_tts(request: TTSStreamRequest):
     if voice not in BUILTIN_VOICES:
         voice = DEFAULT_VOICE
     
-    async def audio_generator():
+    def sync_audio_generator():
-        """Generate audio chunks (async)"""
+        """Generate audio chunks (sync generator)"""
         try:
             syn_tokens = model.generate_speech(
                 prompt=request.text,
                 voice=voice,
             )
             
-            async for audio_chunk in syn_tokens:
+            for audio_chunk in syn_tokens:
                 yield audio_chunk
                 
         except Exception as e:
@@ -544,7 +549,7 @@ async def stream_tts(request: TTSStreamRequest):
             raise
     
     return StreamingResponse(
-        audio_generator(),
+        sync_audio_generator(),
         media_type="audio/wav"
     )