diff --git a/paddlespeech/t2s/modules/transformer/attention.py b/paddlespeech/t2s/modules/transformer/attention.py
index d7a032445..e3c9a992a 100644
--- a/paddlespeech/t2s/modules/transformer/attention.py
+++ b/paddlespeech/t2s/modules/transformer/attention.py
@@ -196,7 +196,7 @@ class RelPositionMultiHeadedAttention(MultiHeadedAttention):
 
         if self.zero_triu:
             ones = paddle.ones((t1, t2))
-            x = x * paddle.tril(ones, t2 - 1)[None, None, :, :]
+            x = x * paddle.tril(ones, t2 - t1)[None, None, :, :]
 
         return x
 
@@ -299,7 +299,7 @@ class LegacyRelPositionMultiHeadedAttention(MultiHeadedAttention):
 
         if self.zero_triu:
             ones = paddle.ones((t1, t2))
-            x = x * paddle.tril(ones, t2 - 1)[None, None, :, :]
+            x = x * paddle.tril(ones, t2 - t1)[None, None, :, :]
 
         return x