diff --git a/demo/web/app.py b/demo/web/app.py
index 1b45117e..f29241c2 100644
--- a/demo/web/app.py
+++ b/demo/web/app.py
@@ -332,17 +332,16 @@ def chunk_to_pcm16(self, chunk: np.ndarray) -> bytes:
         return pcm.tobytes()
 
 
-app = FastAPI()
+from contextlib import asynccontextmanager
 
-
-@app.on_event("startup")
-async def _startup() -> None:
+@asynccontextmanager
+async def lifespan(app: FastAPI):
     model_path = os.environ.get("MODEL_PATH")
     if not model_path:
         raise RuntimeError("MODEL_PATH not set in environment")
 
     device = os.environ.get("MODEL_DEVICE", "cuda")
-    
+
     service = StreamingTTSService(
         model_path=model_path,
         device=device
@@ -354,6 +353,10 @@ async def _startup() -> None:
     app.state.device = device
     app.state.websocket_lock = asyncio.Lock()
     print("[startup] Model ready.")
+    yield
+
+
+app = FastAPI(lifespan=lifespan)
 
 
 def streaming_tts(text: str, **kwargs) -> Iterator[np.ndarray]: