############################
INSTALLATION AND COMPILATION
############################


Uncompress and untar the provided file into a directory:

DIRWORK=$HOME/work
cd $DIRWORK
tar zxf upc_vc_lsf_toolkit.tgz

It creates the directory "$DIRWORK/upc_vc_lsf_toolkit" ($VCDIR from now on) and several subdirectories.

Go to "$VCDIR/prj" and compile:

make release

If you need to debug an unoptimized version of the programs, compile without optimization and with debugging symbols:

make debug



##########################
VC TRAINING
##########################


$VCDIR/data/scripts

Several scripts to train the VC parameters (GMMs, CARTs and residual
database) and to transform a test set of input speech files.

pairs: pairs of source-target speakers to train the system (separate
with tabs)

speaker.stats: mean and std of the logF0 for each speaker in
"pairs". The format of the file is: "mean_$speaker" tab
mean_of_speaker (same with std). All the speakers in "pairs" need to
be included here.
	    
tree_struct.desc: description of the CARTs

Scripts:

0.create_train_validation

Create a training/validation set given the whole development set (file
list) and the percentage of files to be used for training.

1.prepare_speakers

Parametrize the input speech files using LSFs and compute the
residual.

2.train_residual_prediction

Create the residual database for prediction. All the residual voiced
frames resulting from the LSF parametrization of the target speaker
are joint in a single file. A pitch normalization step is applied to
eliminate too short or too long pitch perdios. The source and target
frames are also aligned.

3.train_vocal_tract

Align the source and target data, allowing repetitions in both
sets. Create CART data from the resulting PRM files. Estimate a CART
and train GMMs for each leaf. Compute classification results with the
validation set.

convert_eval

Transform the input test set, using the models obtained before. 

$VCDIR/data/database/sig

Required input files:

Audio files (wav, 16kHz, 16bits)
Pitch files (only time instants, one per line)
Segmentation files (start::middle::end::phoneme)

$VCDIR/data/filelists

List of files for each speaker, named as: 
     spkname.dev.gui

Optional:
     spkname.dev.trn.gui  -  Training set
     spkname.dev.val.gui  -  Validation set
(can be automatically generated using the script 0.create_train_validation)




########################## 
EXAMPLE
##########################

$VCDIR/data/scripts/pairs:

spk01	spk02


$VCDIR/data/scripts/speaker.stats:

mean_spk01      2.0899
std_spk01       0.0607

mean_spk02      2.0414
std_spk02       0.0682


$VCDIR/data/filelists/spk01.dev.gui

spk01_0000
spk01_0007
spk01_0009
...


$VCDIR/data/filelists/spk02.dev.gui

spk02_0000
spk02_0007
spk02_0009
...


$VCDIR/data/database/sig/spk01_0000.pit
$VCDIR/data/database/sig/spk01_0000.txt
$VCDIR/data/database/sig/spk01_0000.wav
$VCDIR/data/database/sig/spk01_0007.pit
$VCDIR/data/database/sig/spk01_0007.txt
$VCDIR/data/database/sig/spk01_0007.wav
...

(same for spk02)


$VCDIR/data/database/sig/spk01_0000.pit:

0.999625
1.00394
1.0095
1.01494
1.02025
...


$VCDIR/data/database/sig/spk01_0000.txt:

0.000000::0.478000::0.956000::_
0.956000::0.960000::0.976000:CL
0.976000::0.992000::1.008000::k
1.008000::1.016000::1.056000::'o
1.056000::1.092000::1.140000::m
...



Training:

cd $VCDIR/data/scripts
./0.create_train_validation 80 ../filelists/spk0{1,2}.dev.gui
./1.prepare_speakers spk01 spk02
./2.train_residual_prediction spk01 spk02
./3.train_vocal_tract spk01 spk02

Conversion:

./convert_eval spk01 spk02 ../filelists/spk01.tst.gui