diff --git a/dnn/lpcnet.py b/dnn/lpcnet.py
index 45a11e91..c54942b3 100644
--- a/dnn/lpcnet.py
+++ b/dnn/lpcnet.py
@@ -11,7 +11,7 @@ import numpy as np
 import h5py
 import sys
 
-rnn_units1=256
+rnn_units1=512
 rnn_units2=32
 pcm_bits = 8
 embed_size = 128
diff --git a/dnn/train_wavenet_audio.py b/dnn/train_wavenet_audio.py
index 63cac35a..a1fbbafd 100755
--- a/dnn/train_wavenet_audio.py
+++ b/dnn/train_wavenet_audio.py
@@ -111,6 +111,7 @@ in_exc = in_exc.astype('uint8')
 
 features = np.reshape(features, (nb_frames, feature_chunk_size, nb_features))
 features = features[:, :, :nb_used_features]
+features[:,:,18:36] = 0
 pred = np.reshape(pred, (nb_frames, pcm_chunk_size, 1))
 pred = pred.astype('uint8')
 
@@ -119,8 +120,8 @@ periods = (50*features[:,:,36:37]+100).astype('int16')
 in_data = np.concatenate([in_data, pred], axis=-1)
 
 # dump models to disk as we go
-checkpoint = ModelCheckpoint('wavenet5p0_{epoch:02d}.h5')
+checkpoint = ModelCheckpoint('lpcnet5_512_10_G32np_{epoch:02d}.h5')
 
 #model.load_weights('wavenet4f2_30.h5')
-model.compile(optimizer=Adam(0.001, amsgrad=True, decay=5e-5), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
-model.fit([in_data, in_exc, features, periods], out_data, batch_size=batch_size, epochs=60, validation_split=0.2, callbacks=[checkpoint, lpcnet.Sparsify(1000, 20000, 200, 0.25)])
+model.compile(optimizer=Adam(0.0005, amsgrad=True, decay=5e-5), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
+model.fit([in_data, in_exc, features, periods], out_data, batch_size=batch_size, epochs=60, validation_split=0.2, callbacks=[checkpoint, lpcnet.Sparsify(1000, 20000, 200, 0.1)])