shithub: opus

--- a/dnn/dump_lpcnet.py

+++ b/dnn/dump_lpcnet.py

@@ -240,7 +240,6 @@

 dump_embedding_layer_impl('gru_a_embed_sig', np.dot(E, W), f, hf)

 W = model.get_layer('gru_a').get_weights()[0][embed_size:2*embed_size,:]

 dump_embedding_layer_impl('gru_a_embed_pred', np.dot(E, W), f, hf)

-E = model.get_layer('embed_exc').get_weights()[0]

 W = model.get_layer('gru_a').get_weights()[0][2*embed_size:3*embed_size,:]

 dump_embedding_layer_impl('gru_a_embed_exc', np.dot(E, W), f, hf)

 W = model.get_layer('gru_a').get_weights()[0][3*embed_size:,:]

--- a/dnn/lpcnet.py

+++ b/dnn/lpcnet.py

@@ -114,8 +114,7 @@

 def new_lpcnet_model(rnn_units1=384, rnn_units2=16, nb_used_features = 38, use_gpu=True):

-    pcm = Input(shape=(None, 2))

-    exc = Input(shape=(None, 1))

+    pcm = Input(shape=(None, 3))

     feat = Input(shape=(None, nb_used_features))

     pitch = Input(shape=(None, 1))

     dec_feat = Input(shape=(None, 128))

@@ -126,9 +125,7 @@

     fconv2 = Conv1D(128, 3, padding='same', activation='tanh', name='feature_conv2')

     embed = Embedding(256, embed_size, embeddings_initializer=PCMInit(), name='embed_sig')

-    cpcm = Reshape((-1, embed_size*2))(embed(pcm))

-    embed2 = Embedding(256, embed_size, embeddings_initializer=PCMInit(), name='embed_exc')

-    cexc = Reshape((-1, embed_size))(embed2(exc))

+    cpcm = Reshape((-1, embed_size*3))(embed(pcm))

     pembed = Embedding(256, 64, name='embed_pitch')

     cat_feat = Concatenate()([feat, Reshape((-1, 64))(pembed(pitch))])

@@ -149,13 +146,13 @@

         rnn = GRU(rnn_units1, return_sequences=True, return_state=True, recurrent_activation="sigmoid", reset_after='true', name='gru_a')

         rnn2 = GRU(rnn_units2, return_sequences=True, return_state=True, recurrent_activation="sigmoid", reset_after='true', name='gru_b')

-    rnn_in = Concatenate()([cpcm, cexc, rep(cfeat)])

+    rnn_in = Concatenate()([cpcm, rep(cfeat)])

     md = MDense(pcm_levels, activation='softmax', name='dual_fc')

     gru_out1, _ = rnn(rnn_in)

     gru_out2, _ = rnn2(Concatenate()([gru_out1, rep(cfeat)]))

     ulaw_prob = md(gru_out2)

-    model = Model([pcm, exc, feat, pitch], ulaw_prob)

+    model = Model([pcm, feat, pitch], ulaw_prob)

     model.rnn_units1 = rnn_units1

     model.rnn_units2 = rnn_units2

     model.nb_used_features = nb_used_features

@@ -163,10 +160,10 @@

     encoder = Model([feat, pitch], cfeat)

-    dec_rnn_in = Concatenate()([cpcm, cexc, dec_feat])

+    dec_rnn_in = Concatenate()([cpcm, dec_feat])

     dec_gru_out1, state1 = rnn(dec_rnn_in, initial_state=dec_state1)

     dec_gru_out2, state2 = rnn2(Concatenate()([dec_gru_out1, dec_feat]), initial_state=dec_state2)

     dec_ulaw_prob = md(dec_gru_out2)

-    decoder = Model([pcm, exc, dec_feat, dec_state1, dec_state2], [dec_ulaw_prob, state1, state2])

+    decoder = Model([pcm, dec_feat, dec_state1, dec_state2], [dec_ulaw_prob, state1, state2])

     return model, encoder, decoder

--- a/dnn/test_lpcnet.py

+++ b/dnn/test_lpcnet.py

@@ -63,13 +63,12 @@

-model.load_weights('lpcnet20c_384_10_G16_80.h5')

+model.load_weights('lpcnet20g_384_10_G16_02.h5')

 order = 16

 pcm = np.zeros((nb_frames*pcm_chunk_size, ))

-fexc = np.zeros((1, 1, 2), dtype='float32')

-iexc = np.zeros((1, 1, 1), dtype='int16')

+fexc = np.zeros((1, 1, 3), dtype='int16')

 state1 = np.zeros((1, model.rnn_units1), dtype='float32')

 state2 = np.zeros((1, model.rnn_units2), dtype='float32')

@@ -88,7 +87,7 @@

             pred = -sum(a*pcm[f*frame_size + i - 1:f*frame_size + i - order-1:-1])

             fexc[0, 0, 1] = lin2ulaw(pred)

-            p, state1, state2 = dec.predict([fexc, iexc, cfeat[:, fr:fr+1, :], state1, state2])

+            p, state1, state2 = dec.predict([fexc, cfeat[:, fr:fr+1, :], state1, state2])

             #Lower the temperature for voiced frames to reduce noisiness

             p *= np.power(p, np.maximum(0, 1.5*features[c, fr, 37] - .5))

             p = p/(1e-18 + np.sum(p))

@@ -96,8 +95,8 @@

             p = np.maximum(p-0.002, 0).astype('float64')

             p = p/(1e-8 + np.sum(p))

-            iexc[0, 0, 0] = np.argmax(np.random.multinomial(1, p[0,0,:], 1))

-            pcm[f*frame_size + i] = pred + ulaw2lin(iexc[0, 0, 0])

+            fexc[0, 0, 2] = np.argmax(np.random.multinomial(1, p[0,0,:], 1))

+            pcm[f*frame_size + i] = pred + ulaw2lin(fexc[0, 0, 2])

             fexc[0, 0, 0] = lin2ulaw(pcm[f*frame_size + i])

             mem = coef*mem + pcm[f*frame_size + i]

             #print(mem)

--- a/dnn/train_lpcnet.py

+++ b/dnn/train_lpcnet.py

@@ -91,10 +91,11 @@

 periods = (.1 + 50*features[:,:,36:37]+100).astype('int16')

-in_data = np.concatenate([sig, pred], axis=-1)

+in_data = np.concatenate([sig, pred, in_exc], axis=-1)

 del sig

 del pred

+del in_exc

 # dump models to disk as we go

 checkpoint = ModelCheckpoint('lpcnet20g_384_10_G16_{epoch:02d}.h5')

@@ -101,4 +102,4 @@

 #model.load_weights('lpcnet9b_384_10_G16_01.h5')

 model.compile(optimizer=Adam(0.001, amsgrad=True, decay=5e-5), loss='sparse_categorical_crossentropy')

-model.fit([in_data, in_exc, features, periods], out_exc, batch_size=batch_size, epochs=nb_epochs, validation_split=0.0, callbacks=[checkpoint, lpcnet.Sparsify(2000, 40000, 400, (0.05, 0.05, 0.2))])

+model.fit([in_data, features, periods], out_exc, batch_size=batch_size, epochs=nb_epochs, validation_split=0.0, callbacks=[checkpoint, lpcnet.Sparsify(2000, 40000, 400, (0.05, 0.05, 0.2))])

--

⑨