diff --git a/paddlespeech/t2s/modules/tacotron2/attentions.py b/paddlespeech/t2s/modules/tacotron2/attentions.py
index cdaef4608..2256d8ea8 100644
--- a/paddlespeech/t2s/modules/tacotron2/attentions.py
+++ b/paddlespeech/t2s/modules/tacotron2/attentions.py
@@ -47,11 +47,13 @@ def _apply_attention_constraint(e,
         https://arxiv.org/abs/1710.07654
 
     """
-    if paddle.shape(e)[0] != 1:
-        raise NotImplementedError(
-            "Batch attention constraining is not yet supported.")
-    backward_idx = last_attended_idx - backward_window
-    forward_idx = last_attended_idx + forward_window
+    # for dygraph to static graph
+    # if e.shape[0] != 1:
+    #     raise NotImplementedError(
+    #         "Batch attention constraining is not yet supported.")
+    backward_idx = paddle.cast(
+        last_attended_idx - backward_window, dtype='int64')
+    forward_idx = paddle.cast(last_attended_idx + forward_window, dtype='int64')
     if backward_idx > 0:
         e[:, :backward_idx] = -float("inf")
     if forward_idx < paddle.shape(e)[1]:
diff --git a/paddlespeech/t2s/modules/tacotron2/decoder.py b/paddlespeech/t2s/modules/tacotron2/decoder.py
index 41c94b63f..6118a004e 100644
--- a/paddlespeech/t2s/modules/tacotron2/decoder.py
+++ b/paddlespeech/t2s/modules/tacotron2/decoder.py
@@ -562,7 +562,7 @@ class Decoder(nn.Layer):
         idx = 0
         outs, att_ws, probs = [], [], []
         prob = paddle.zeros([1])
-        while True:
+        while paddle.to_tensor(True):
             # updated index
             idx += self.reduction_factor