good results in 200406a, some issues with pitch not changing in females

src/700c_train.sh

@@ -16,7 +16,7 @@ train2=train_8k
 test1=all_8k
 test2=all_speech_subset_8k
 datestamp=$1
-epochs=10
+epochs=05
 log=${1}.txt
 train=${datestamp}_train
 
@@ -33,15 +33,11 @@ experiment() {
     echo "train starting" ${2}
     echo "------------------------------------------------------------------------------"
     
-    sox -r 8000 -c 1 ~/Downloads/${train1}.sw \
-	-r 8000 -c 1 ~/Downloads/${train2}.sw \
-	-t sw -r 8000 -c 1 ${train}.sw
-    
     c2sim ${train}.sw --ten_ms_centre ${train}_10ms.sw --rateKWov ${train}.f32 ${1}
     sw2packedulaw --frame_size 80 ${train}_10ms.sw ${train}.f32 ${train}_10ms.pulaw
     
-    train_lpcnet.py ${train}.f32 ${train}_10ms.pulaw ${datestamp} --epochs ${epochs} --frame_size 80
-    dump_lpcnet.py ${datestamp}_${epochs}.h5
+    train_lpcnet.py ${train}.f32 ${train}_10ms.pulaw ${datestamp}_${2} --epochs ${epochs} --frame_size 80
+    dump_lpcnet.py ${datestamp}_${2}_${epochs}.h5
     cp nnet_data.c src
     make test_lpcnet
 
@@ -51,8 +47,11 @@ experiment() {
 
 rm -f $log
 
-# Quantised 700C vectors at 10ms frame rate (note LPCs unquantised)
-
+# assemble some training speech
+sox -r 8000 -c 1 ~/Downloads/${train1}.sw \
+    -r 8000 -c 1 ~/Downloads/${train2}.sw \
+    -t sw -r 8000 -c 1 ${train}.sw
+    
 (
     experiment "" "none"           # no prediction
 

src/lpcnet.c

@@ -184,9 +184,7 @@ void lpcnet_synthesize(LPCNetState *lpcnet, short *output, float *features, int
     case 2:
         for (i=0;i<LPC_ORDER;i++) {
 	    lpcnet->old_lpc[0][i] = features[i+NB_BANDS];
-	    fprintf(stderr, "%f ", lpcnet->old_lpc[0][i]);
 	}
-	fprintf(stderr, "\n");
 	break;
     default:
 	assert(0);

src/train_lpcnet.py

@@ -155,4 +155,4 @@ checkpoint = ModelCheckpoint(prefix + '_{epoch:02d}.h5')
 # use this to reload a partially trained model
 #model.load_weights('lpcnet_190203_07.h5')
 model.compile(optimizer=Adam(0.001, amsgrad=True, decay=5e-5), loss='sparse_categorical_crossentropy')
-model.fit([in_data, features, periods], out_exc, batch_size=batch_size, epochs=nb_epochs, validation_split=0.1, 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, callbacks=[checkpoint, lpcnet.Sparsify(2000, 40000, 400, (0.05, 0.05, 0.2))])