diff --git a/paddlespeech/t2s/modules/tacotron2/attentions.py b/paddlespeech/t2s/modules/tacotron2/attentions.py
index 2256d8ea8..691139f10 100644
--- a/paddlespeech/t2s/modules/tacotron2/attentions.py
+++ b/paddlespeech/t2s/modules/tacotron2/attentions.py
@@ -194,7 +194,7 @@ class AttLoc(nn.Layer):
 
         e = masked_fill(e, self.mask, -float("inf"))
         # apply monotonic attention constraint (mainly for TTS)
-        if last_attended_idx is not None:
+        if last_attended_idx != -1:
             e = _apply_attention_constraint(e, last_attended_idx,
                                             backward_window, forward_window)
 
diff --git a/paddlespeech/t2s/modules/tacotron2/decoder.py b/paddlespeech/t2s/modules/tacotron2/decoder.py
index 6118a004e..15e29194d 100644
--- a/paddlespeech/t2s/modules/tacotron2/decoder.py
+++ b/paddlespeech/t2s/modules/tacotron2/decoder.py
@@ -556,13 +556,15 @@ class Decoder(nn.Layer):
         if use_att_constraint:
             last_attended_idx = 0
         else:
-            last_attended_idx = None
+            last_attended_idx = -1
 
         # loop for an output sequence
         idx = 0
         outs, att_ws, probs = [], [], []
         prob = paddle.zeros([1])
         while paddle.to_tensor(True):
+            z_list = z_list
+            c_list = c_list
             # updated index
             idx += self.reduction_factor