diff --git a/dnn/lpcnet.py b/dnn/lpcnet.py
index 85fa3da6..7d40c1a5 100644
--- a/dnn/lpcnet.py
+++ b/dnn/lpcnet.py
@@ -41,7 +41,7 @@ class PCMInit(Initializer):
         }
 
 def new_wavernn_model():
-    pcm = Input(shape=(None, 1))
+    pcm = Input(shape=(None, 2))
     pitch = Input(shape=(None, 1))
     feat = Input(shape=(None, nb_used_features))
     dec_feat = Input(shape=(None, 32))
@@ -61,7 +61,7 @@ def new_wavernn_model():
         cpitch = pitch
 
     embed = Embedding(256, 128, embeddings_initializer=PCMInit())
-    cpcm = Reshape((-1, 128))(embed(pcm))
+    cpcm = Reshape((-1, 128*2))(embed(pcm))
 
 
     cfeat = fconv2(fconv1(feat))
diff --git a/dnn/train_wavenet_audio.py b/dnn/train_wavenet_audio.py
index b6c46c08..df8c39cb 100755
--- a/dnn/train_wavenet_audio.py
+++ b/dnn/train_wavenet_audio.py
@@ -51,8 +51,18 @@ features = np.reshape(features, (nb_frames*feature_chunk_size, nb_features))
 
 pred = np.fromfile(pred_file, dtype='int16')
 pred = pred[:nb_frames*pcm_chunk_size]
+
+pred_in = 32768.*ulaw2lin(data)
+for i in range(2, nb_frames*feature_chunk_size):
+    pred[i*frame_size:(i+1)*frame_size] = 0
+    if i % 100000 == 0:
+        print(i)
+    for k in range(16):
+        pred[i*frame_size:(i+1)*frame_size] = pred[i*frame_size:(i+1)*frame_size] - \
+            pred_in[i*frame_size-k-1:(i+1)*frame_size-k-1]*features[i, nb_features-16+k]
+
 pred = np.minimum(127, lin2ulaw(pred/32768.))
-pred = pred + np.random.randint(-1, 1, len(data))
+#pred = pred + np.random.randint(-1, 1, len(data))
 
 
 pitch = 1.*data
@@ -72,7 +82,7 @@ features = features[:, :, :nb_used_features]
 pred = np.reshape(pred, (nb_frames, pcm_chunk_size, 1))
 pred = (pred.astype('int16')+128).astype('uint8')
 
-#in_data = np.concatenate([in_data, pred], axis=-1)
+in_data = np.concatenate([in_data, pred], axis=-1)
 
 #in_data = np.concatenate([in_data, in_pitch], axis=-1)