diff --git a/dnn/denoise.c b/dnn/denoise.c
index dcb225bb..84c7d7ea 100644
--- a/dnn/denoise.c
+++ b/dnn/denoise.c
@@ -577,6 +577,7 @@ int main(int argc, char **argv) {
       return 0;
     }
     for (i=0;i<FRAME_SIZE;i++) x[i] = tmp[i];
+    for (i=0;i<FRAME_SIZE;i++) x[i] += rand()/(float)RAND_MAX - .5;
     for (i=0;i<FRAME_SIZE;i++) E += tmp[i]*(float)tmp[i];
     biquad(x, mem_hp_x, x, b_hp, a_hp, FRAME_SIZE);
     preemphasis(x, &mem_preemph, x, PREEMPHASIS, FRAME_SIZE);
diff --git a/dnn/lpcnet.py b/dnn/lpcnet.py
index dcb7f3dd..5072c589 100644
--- a/dnn/lpcnet.py
+++ b/dnn/lpcnet.py
@@ -12,7 +12,7 @@ import sys
 rnn_units=512
 pcm_bits = 8
 pcm_levels = 2**pcm_bits
-nb_used_features = 37
+nb_used_features = 38
 
 
 def new_wavernn_model():
@@ -22,11 +22,11 @@ def new_wavernn_model():
     dec_feat = Input(shape=(None, 32))
     dec_state = Input(shape=(rnn_units,))
 
-    conv1 = Conv1D(16, 7, padding='causal')
-    pconv1 = Conv1D(16, 5, padding='same')
-    pconv2 = Conv1D(16, 5, padding='same')
-    fconv1 = Conv1D(128, 3, padding='same')
-    fconv2 = Conv1D(32, 3, padding='same')
+    conv1 = Conv1D(16, 7, padding='causal', activation='tanh')
+    pconv1 = Conv1D(16, 5, padding='same', activation='tanh')
+    pconv2 = Conv1D(16, 5, padding='same', activation='tanh')
+    fconv1 = Conv1D(128, 3, padding='same', activation='tanh')
+    fconv2 = Conv1D(32, 3, padding='same', activation='tanh')
 
     if False:
         cpcm = conv1(pcm)
@@ -40,17 +40,17 @@ def new_wavernn_model():
     rep = Lambda(lambda x: K.repeat_elements(x, 160, 1))
 
     rnn = CuDNNGRU(rnn_units, return_sequences=True, return_state=True)
-    rnn_in = Concatenate()([cpcm, cpitch, rep(cfeat)])
+    rnn_in = Concatenate()([cpcm, rep(cfeat)])
     md = MDense(pcm_levels, activation='softmax')
     gru_out, state = rnn(rnn_in)
     ulaw_prob = md(gru_out)
     
-    model = Model([pcm, pitch, feat], ulaw_prob)
+    model = Model([pcm, feat], ulaw_prob)
     encoder = Model(feat, cfeat)
     
-    dec_rnn_in = Concatenate()([cpcm, cpitch, dec_feat])
+    dec_rnn_in = Concatenate()([cpcm, dec_feat])
     dec_gru_out, state = rnn(dec_rnn_in, initial_state=dec_state)
     dec_ulaw_prob = md(dec_gru_out)
 
-    decoder = Model([pcm, pitch, dec_feat, dec_state], [dec_ulaw_prob, state])
+    decoder = Model([pcm, dec_feat, dec_state], [dec_ulaw_prob, state])
     return model, encoder, decoder
diff --git a/dnn/test_lpcnet.py b/dnn/test_lpcnet.py
index 4c47274e..5b194554 100755
--- a/dnn/test_lpcnet.py
+++ b/dnn/test_lpcnet.py
@@ -47,7 +47,7 @@ in_data = np.reshape(in_data, (nb_frames*pcm_chunk_size, 1))
 out_data = np.reshape(data, (nb_frames*pcm_chunk_size, 1))
 
 
-model.load_weights('lpcnet1i_30.h5')
+model.load_weights('lpcnet3a_21.h5')
 
 order = 16
 
@@ -61,7 +61,7 @@ for c in range(1, nb_frames):
     cfeat = enc.predict(features[c:c+1, :, :nb_used_features])
     for fr in range(1, feature_chunk_size):
         f = c*feature_chunk_size + fr
-        a = features[c, fr, nb_used_features+1:]
+        a = features[c, fr, nb_used_features:]
         
         #print(a)
         gain = 1.;
@@ -69,9 +69,10 @@ for c in range(1, nb_frames):
         period = period - 4
         for i in range(frame_size):
             pitch[0, 0, 0] = exc[f*frame_size + i - period, 0]
-            fexc[0, 0, 0] = exc[f*frame_size + i - 1]
+            fexc[0, 0, 0] = 2*exc[f*frame_size + i - 1]
+            #fexc[0, 0, 0] = in_data[f*frame_size + i, 0]
             #print(cfeat.shape)
-            p, state = dec.predict([fexc, pitch, cfeat[:, fr:fr+1, :], state])
+            p, state = dec.predict([fexc, cfeat[:, fr:fr+1, :], state])
             p = p/(1e-5 + np.sum(p))
             #print(np.sum(p))
             iexc[0, 0, 0] = np.argmax(np.random.multinomial(1, p[0,0,:], 1))-128
diff --git a/dnn/train_lpcnet.py b/dnn/train_lpcnet.py
index 833dc34c..8658b01e 100755
--- a/dnn/train_lpcnet.py
+++ b/dnn/train_lpcnet.py
@@ -13,13 +13,13 @@ from adadiff import Adadiff
 import tensorflow as tf
 from keras.backend.tensorflow_backend import set_session
 config = tf.ConfigProto()
-config.gpu_options.per_process_gpu_memory_fraction = 0.28
+config.gpu_options.per_process_gpu_memory_fraction = 0.44
 set_session(tf.Session(config=config))
 
 nb_epochs = 40
 batch_size = 64
 
-model = lpcnet.new_wavernn_model()
+model, enc, dec = lpcnet.new_wavernn_model()
 model.compile(optimizer=Adadiff(), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
 model.summary()
 
@@ -63,8 +63,8 @@ features = features[:, :, :nb_used_features]
 # f.create_dataset('data', data=in_data[:50000, :, :])
 # f.create_dataset('feat', data=features[:50000, :, :])
 
-checkpoint = ModelCheckpoint('lpcnet1k_{epoch:02d}.h5')
+checkpoint = ModelCheckpoint('lpcnet3b_{epoch:02d}.h5')
 
 #model.load_weights('wavernn1c_01.h5')
-model.compile(optimizer=Adadiff(), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
-model.fit([in_data, in_pitch, features], out_data, batch_size=batch_size, epochs=30, validation_split=0.2, callbacks=[checkpoint])
+model.compile(optimizer=Adam(0.001, amsgrad=True, decay=2e-4), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
+model.fit([in_data, features], out_data, batch_size=batch_size, epochs=30, validation_split=0.2, callbacks=[checkpoint])