190920c in between 190215 and 190920a, not quite good enough

src/full_train.sh

@@ -0,0 +1,48 @@
+#!/bin/bash -x
+# full_train.sh
+# Script to train using full database, maerial developed to give reasonable quality
+# to test sample database.
+#
+# usage:
+#   $ cd LPCNet/build_linux
+#   $ ../src/full_train.sh
+
+SRC1=david_16kHz.wav             # 122s
+SRC2=vk5apr_recording_21_may.wav # 64s
+SRC3=all_speechcat.sw            # 185 minutes, wide range of speakers
+SRC4=wianews-2019-01-20.s16      # 62 minutes
+SRC5=soldersmoke202.sw           # 76 minutes, lot of material with target voices
+SRC=train_src
+
+DATE=190920c
+
+synth() {
+  ./src/dump_data --test --c2pitch --c2voicing ~/Downloads/$1.sw $1.f32
+  ./src/test_lpcnet $1.f32 "$2".raw
+}
+
+train() {
+    # repeat David and Peter to get 60 minutes worth, so it weights training
+    # for this type of speaker that we are struggling with
+    x=$(mktemp)
+    sox ~/Downloads/$SRC1 ~/Downloads/$SRC2 $x'.wav' repeat 20
+    ls -l $x.wav
+    # combine all samples, evaluation data at end of larger database of mixed speakers
+    sox $x.wav \
+	-t sw -r 16000 -c 1 ~/Downloads/$SRC4 \
+        -t sw -r 16000 -c 1 ~/Downloads/$SRC5 \
+	-t sw -r 16000 -c 1 ~/Downloads/$SRC3 \
+	-t sw $SRC.sw
+    ls -l $SRC.sw
+    ./src/dump_data --train --c2pitch --c2voicing -z 1 $SRC.sw $SRC.f32 $SRC.pcm
+    ../src/train_lpcnet.py $SRC.f32 $SRC.pcm lpcnet_$DATE
+    ../src/dump_lpcnet.py lpcnet_"$DATE"_10.h5
+    cp nnet_data.c src
+    make test_lpcnet
+}
+
+train
+synth c01_01 $DATE'_f'
+synth mk61_01 $DATE'_m'
+synth cq_16kHz $DATE'_cq_16kHz'
+synth peter $DATE'_peter'

src/peter_david_train.sh

@@ -13,7 +13,7 @@ SRC1=david_16kHz.wav             # 122s
 SRC2=vk5apr_recording_21_may.wav # 64s
 SRC=train_src
 
-DATE=190920a
+DATE=190920b
 
 synth() {
    ./src/dump_data --test --c2pitch --c2voicing ~/Downloads/$1.sw $1.f32 

src/tinytrain.sh

@@ -7,34 +7,36 @@
 #   $ cd LPCNet/build_linux
 #   $ ../src/tinytrain.sh
 
-# TODO:
-# + maybe rname to a separate train script so we can repeat in future
-# + test utterance of Peter and I
-# + this is an unusual test of very small training database, to see if it can
-#   work within this
-
-SRC1=david_16kHz.wav             # 122s
+SRC1=david_16kHz.wav # 122s
 SRC2=vk5apr_recording_21_may.wav # 64s
+SRC3=all_speechcat.sw
+SRC4=wianews-2019-01-20.s16
+SRC5=bob.wav
 SRC=train_src
 
 DATE=190806b
 
 synth() {
-   ./src/dump_data --mag --test --c2pitch --c2voicing ~/Downloads/$1.sw $1.f32 
+  ./src/dump_data --mag --test --c2pitch --c2voicing ~/Downloads/$1.sw $1.f32
   ./src/test_lpcnet --mag $1.f32 "$2".raw
 }
 
 train() {
-    sox ~/Downloads/$SRC1 ~/Downloads/$SRC2 -t sw $SRC.sw
-    ./src/dump_data --mag --train --c2pitch --c2voicing -z 1 -n 1E6 $SRC.sw $SRC.f32 $SRC.pcm
+    sox ~/Downloads/$SRC1 \
+	-r 16000 ~/Downloads/$SRC2 \
+	-t sw -r 16000 ~/Downloads/$SRC3 \
+	-t sw -r 16000 -c 1 ~/Downloads/$SRC4 \
+        ~/Downloads/$SRC5 \
+	-t sw $SRC.sw
+    ./src/dump_data --mag --train --c2pitch --c2voicing -z 1 $SRC.sw $SRC.f32 $SRC.pcm
     ../src/train_lpcnet.py $SRC.f32 $SRC.pcm lpcnet_$DATE
     ../src/dump_lpcnet.py lpcnet_"$DATE"_10.h5
     cp nnet_data.c src
     make test_lpcnet
 }
 
-#train
+train
 #synth c01_01 $DATE'_f'
 #synth mk61_01 $DATE'_m'
-#synth wia $DATE'_wia'
 synth bob $DATE'_bob'
+synth wia $DATE'_wia'

src/train_lpcnet.py

@@ -46,7 +46,7 @@ config = tf.ConfigProto()
 
 set_session(tf.Session(config=config))
 
-nb_epochs = 10
+nb_epochs = 20
 
 # Try reducing batch_size if you run out of memory on your GPU
 batch_size = 32