diff --git a/paddlespeech/t2s/models/fastspeech2/fastspeech2.py b/paddlespeech/t2s/models/fastspeech2/fastspeech2.py
index fcd54f0d2..91bfc540a 100644
--- a/paddlespeech/t2s/models/fastspeech2/fastspeech2.py
+++ b/paddlespeech/t2s/models/fastspeech2/fastspeech2.py
@@ -903,14 +903,14 @@ class FastSpeech2(nn.Layer):
 
         # initialize alpha in scaled positional encoding
         if self.encoder_type == "transformer" and self.use_scaled_pos_enc:
-            init_enc_alpha = paddle.to_tensor(init_enc_alpha)
+            init_enc_alpha = paddle.to_tensor(init_enc_alpha).reshape([1])
             self.encoder.embed[-1].alpha = paddle.create_parameter(
                 shape=init_enc_alpha.shape,
                 dtype=str(init_enc_alpha.numpy().dtype),
                 default_initializer=paddle.nn.initializer.Assign(
                     init_enc_alpha))
         if self.decoder_type == "transformer" and self.use_scaled_pos_enc:
-            init_dec_alpha = paddle.to_tensor(init_dec_alpha)
+            init_dec_alpha = paddle.to_tensor(init_dec_alpha).reshape([1])
             self.decoder.embed[-1].alpha = paddle.create_parameter(
                 shape=init_dec_alpha.shape,
                 dtype=str(init_dec_alpha.numpy().dtype),