fix paddling len bug

deepspeech/exps/deepspeech2/model.py

@@ -455,10 +455,12 @@ class DeepSpeech2ExportTester(DeepSpeech2Tester):
         x_batch = audio.numpy()
         batch_size, Tmax, x_dim = x_batch.shape
         x_len_batch = audio_len.numpy().astype(np.int64)
-
-        padding_len_batch = chunk_stride - (
-            Tmax - chunk_size
-        ) % chunk_stride  # The length of padding for the batch
+        if (Tmax - chunk_size) % chunk_stride != 0:
+            padding_len_batch = chunk_stride - (
+                Tmax - chunk_size
+            ) % chunk_stride  # The length of padding for the batch
+        else:
+            padding_len_batch = 0
         x_list = np.split(x_batch, batch_size, axis=0)
         x_len_list = np.split(x_len_batch, batch_size, axis=0)
 

deepspeech/models/ds2_online/deepspeech2.py

@@ -100,12 +100,12 @@ class CRNNEncoder(nn.Layer):
         """Compute Encoder outputs
 
         Args:
-            x (Tensor): [B, feature_size, D]
+            x (Tensor): [B, T, D]
             x_lens (Tensor): [B]
             init_state_h_box(Tensor): init_states h for RNN layers: [num_rnn_layers * num_directions, batch_size, hidden_size]
             init_state_c_box(Tensor): init_states c for RNN layers: [num_rnn_layers * num_directions, batch_size, hidden_size]
         Return:
-            x (Tensor): encoder outputs, [B, size, D]
+            x (Tensor): encoder outputs, [B, T, D]
             x_lens (Tensor): encoder length, [B]
             final_state_h_box(Tensor): final_states h for RNN layers: [num_rnn_layers * num_directions, batch_size, hidden_size]
             final_state_c_box(Tensor): final_states c for RNN layers: [num_rnn_layers * num_directions, batch_size, hidden_size]