diff --git a/whisper/decoding.py b/whisper/decoding.py
index 49485d009..0a5143cc8 100644
--- a/whisper/decoding.py
+++ b/whisper/decoding.py
@@ -655,9 +655,9 @@ def _get_audio_features(self, mel: Tensor):
             audio_features = self.model.encoder(mel)
 
         if audio_features.dtype != (
-            torch.float16 if self.options.fp16 else torch.float32
+                torch.float16 if self.options.fp16 else torch.float32
         ):
-            return TypeError(
+            raise TypeError(
                 f"audio_features has an incorrect dtype: {audio_features.dtype}"
             )