diff --git a/openvoice/se_extractor.py b/openvoice/se_extractor.py
index a41c2566..c3fa01ae 100644
--- a/openvoice/se_extractor.py
+++ b/openvoice/se_extractor.py
@@ -19,7 +19,10 @@
 def split_audio_whisper(audio_path, audio_name, target_dir='processed'):
     global model
     if model is None:
-        model = WhisperModel(model_size, device="cuda", compute_type="float16")
+        if torch.cuda.is_available():
+            model = WhisperModel(model_size, device="cuda", compute_type="float16")
+        else:
+            model = WhisperModel(model_size, device="cpu", compute_type="int8")
     audio = AudioSegment.from_file(audio_path)
     max_len = len(audio)