diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py
index 3de21a293..9fe88929d 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py
@@ -66,19 +66,14 @@ class FrameReducer(nn.Module):
 
         padding_mask = make_pad_mask(x_lens)
         non_blank_mask = (ctc_output[:, :, blank_id] < math.log(0.9)) * (~padding_mask)
-        T_range = torch.arange(x.shape[1], device=x.device)
 
         frames_list: List[torch.Tensor] = []
         lens_list: List[int] = []
         for i in range(x.shape[0]):
-            indexes = torch.masked_select(
-                T_range,
-                non_blank_mask[i],
-            )
-            frames = x[i][indexes]
+            frames = x[i][non_blank_mask[i]]
             frames_list.append(frames)
             lens_list.append(frames.shape[0])
-        x_fr = pad_sequence(frames_list).transpose(0, 1)
+        x_fr = pad_sequence(frames_list, batch_first=True)
         x_lens_fr = torch.tensor(lens_list).to(device=x.device)
 
         return x_fr, x_lens_fr