diff --git a/speech.py b/speech.py
index dbcc4d5..6e6d358 100755
--- a/speech.py
+++ b/speech.py
@@ -20,6 +20,24 @@ from pydantic import BaseModel
 from typing import Optional
 import uvicorn
 
+# Try to import XTTS dependencies (might not be available in minimal installations)
+try:
+    import torch
+    from TTS.tts.configs.xtts_config import XttsConfig
+    from TTS.tts.models.xtts import Xtts
+    from TTS.utils.manage import ModelManager
+    from TTS.tts.layers.xtts.tokenizer import split_sentence
+    from langdetect import detect
+    XTTS_AVAILABLE = True
+except ImportError:
+    XTTS_AVAILABLE = False
+    torch = None
+    XttsConfig = None
+    Xtts = None
+    ModelManager = None
+    split_sentence = None
+    detect = None
+
 @contextlib.asynccontextmanager
 async def lifespan(app):
     # Startup: Initialize voice caches in each worker process
@@ -818,15 +836,9 @@ if __name__ == "__main__":
     logger.remove()
     logger.add(sink=sys.stderr, level=args.log_level)
 
-    if args.xtts_device != "none":
-        import torch
-        from TTS.tts.configs.xtts_config import XttsConfig
-        from TTS.tts.models.xtts import Xtts
-        from TTS.utils.manage import ModelManager
-        from TTS.tts.layers.xtts.tokenizer import split_sentence
-        from langdetect import detect
-
-    if args.preload:
+    if args.preload and not XTTS_AVAILABLE:
+        logger.error("Cannot preload XTTS model - XTTS dependencies not available")
+    elif args.preload:
         xtts = xtts_wrapper(args.preload, device=args.xtts_device, unload_timer=args.unload_timer)
 
     app.register_model('tts-1')