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 c44cb1eaf..b218c9f94 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py
@@ -76,7 +76,7 @@ class FrameReducer(nn.Module):
 
         if y_lens is not None:
             # Limit the maximum number of reduced frames
-            limit_lens = T - y_lens
+            limit_lens = x_lens - y_lens
             max_limit_len = limit_lens.max().int()
             fake_limit_indexes = torch.topk(
                 ctc_output[:, :, blank_id], max_limit_len
@@ -97,6 +97,7 @@ class FrameReducer(nn.Module):
             ).scatter_(1, limit_indexes, 1)
 
             non_blank_mask = non_blank_mask | ~limit_mask
+            non_blank_mask = non_blank_mask & ~padding_mask
 
         out_lens = non_blank_mask.sum(dim=1)
         max_len = out_lens.max()