lmaxwell / gmmVoiceConversion

gmm-based voice conversion

gmm-based voice conversion

Files

1. run.sh basic modules

2. config.sh config file

3. go.sh glue basic modules together

tutorial

just follow go.sh

#!/bin/bash
set -e

#mixture number of gmm
MIX=1
config=config.sh

# your data directory, should include directories of source and target speaker
#for example: /highway2/data/vc/huo /highway2/data/vc/wang
datdir=/highway2/data/vc

# gmm training directory, all files generated in training are stored here.
workdir=/highway2/gmmtraining

echo "===========prepare train and test file================"
# prepare train and test files
# automaticaly random select 1/10 files for test, the rest are for training
./run.sh -f $config -d $datdir -w $workdir -t prepare


echo "===========calculate gv of target singer=============="
./run.sh -f $config -d $datdir -w $workdir -t gv



echo "===========statistics of f0 of source and target singer=============="
./run.sh -f $config -d $datdir -w $workdir -t f0stat


echo "===========do iterative training====================================="
for ite in `seq 0 5`
do
    echo "============iteration $ite===================================="
    echo "============dtw align"
    ./run.sh -f $config -d $datdir -w $workdir -t dtw -m $MIX -i $ite
    echo "============gmm training"
    ./run.sh -f $config -d $datdir -w $workdir -t train -m $MIX -i $ite
    echo "============test"
    ./run.sh -f $config -d $datdir -w $workdir -t test -m $MIX -i $ite
done
#

MIX=64
for i in `seq 5 8`
do
    ./run.sh -f $config -d $datdir  -w $workdir -t train -m $MIX -i $i 
    ./run.sh -f $config -d $datdir  -w $workdir -t test -m $MIX -i $i 
    ./run.sh -f $config -d $datdir  -w $workdir -t dtw -m $MIX -i `echo $i+1|bc` 
done

References

1. Toda T, Black A W, Tokuda K. Voice conversion based on maximum-likelihood estimation of spectral parameter trajectory[J]. Audio, Speech, and Language Processing, IEEE Transactions on, 2007, 15(8): 2222-2235.
2. Kobayashi K, Toda T, Neubig G, et al. Statistical singing voice conversion with direct waveform modification based on the spectrum differential[C]//INTERSPEECH. 2014: 2514-2518.