diff --git a/whisper/decoding.py b/whisper/decoding.py
index 49485d0..8d14dbf 100644
--- a/whisper/decoding.py
+++ b/whisper/decoding.py
@@ -730,7 +730,8 @@ class DecodingTask:
                 )
             ]
 
-        # repeat text tensors by the group size, for beam search or best-of-n sampling
+        # repeat the audio & text tensors by the group size, for beam search or best-of-n sampling
+        audio_features = audio_features.repeat_interleave(self.n_group, dim=0)
         tokens = tokens.repeat_interleave(self.n_group, dim=0).to(audio_features.device)
 
         # call the main sampling loop