Github 项目推荐 | 用 Pytorch 实现的 WaveNet-Vocoder

$ git clone https://github.com/kan-bayashi/PytorchWaveNetVocoder.git
$ cd PytorchWaveNetVocoder/tools
$ make -j

# build SD model
$ cd egs/arctic/sd
$ ./run.sh 

# build SI-CLOSE model
$ cd egs/arctic/si-close
$ ./run.sh 

# build SI-OPEN model
$ cd egs/arctic/si-open
$ ./run.sh

$ cd egs/arctic/sd

# edit configuration
$ vim cmd.sh # please edit as follows
---
# for local
# export train_cmd="run.pl"
# export cuda_cmd="run.pl --gpu 1"

# for slurm (you can change configuration file "conf/slurm.conf")
export train_cmd="slurm.pl --config conf/slurm.conf"
export cuda_cmd="slurm.pl --gpu 1 --config conf/slurm.conf"
---

$ vim conf/slurm.conf # edit <your_partition_name>
---
command sbatch --export=PATH  --ntasks-per-node=1
option time=* --time $0
option mem=* --mem-per-cpu $0
option mem=0
option num_threads=* --cpus-per-task $0 --ntasks-per-node=1
option num_threads=1 --cpus-per-task 1  --ntasks-per-node=1
default gpu=0
option gpu=0 -p <your_partion_name>
option gpu=* -p <your_partion_name> --gres=gpu:$0 --time 10-00:00:00
---

# run the recipe
$ ./run.sh

- checkpoint-final.pkl (model parameter file)
- model.conf (model configuration file)
- stats.h5 (feature statistics file)
- *.wav (your own wav file)

$ cd egs/arctic/si-close

# download pre-trained model which trained with 6 arctic speakers
$ wget "https://www.dropbox.com/s/xt7qqmfgamwpqqg/si-close_lr1e-4_wd0_bs20k_ns_up.zip?dl=0" -O si-close_lr1e-4_wd0_bs20k_ns_up.zip

# unzip 
$ unzip si-close_lr1e-4_wd0_bs20k_ns_up.zip

# make filelist of your own wav files
$ find <your_wav_dir> -name "*.wav" > wav.scp

# feature extraction
$ . ./path.sh
$ feature_extract.py \
    --waveforms wav.scp \
    --wavdir wav/test \
    --hdf5dir hdf5/test \
    --fs 16000 \
    --shiftms 5 \
    --minf0 <set_appropriate_value> \
    --maxf0 <set_appropriate_value> \
    --mcep_dim 24 \
    --mcep_alpha 0.41 \
    --highpass_cutoff 70 \
    --fftl 1024 \
    --n_jobs 1 
    
# make filelist of feature file
$ find hdf5/test -name "*.h5" > feats.scp
    
# decode with pre-trained model
$ decode.py \
    --feats feats.scp \
    --stats si-close_lr1e-4_wd0_bs20k_ns_up/stats.h5 \
    --outdir si-close_lr1e-4_wd0_bs20k_ns_up/wav \
    --checkpoint si-close_lr1e-4_wd0_bs20k_ns_up/checkpoint-final.pkl \
    --config si-close_lr1e-4_wd0_bs20k_ns_up/model.conf \
    --fs 16000 \
    --n_jobs 1 \
    --n_gpus 1

# make filelist of generated wav file
$ find si-close_lr1e-4_wd0_bs20k_ns_up/wav -name "*.wav" > wav_generated.scp

# restore noise shaping
$ noise_shaping.py \
    --waveforms wav_generated.scp \
    --stats si-close_lr1e-4_wd0_bs20k_ns_up/stats.h5 \
    --writedir si-close_lr1e-4_wd0_bs20k_ns_up/wav_restored \
    --fs 16000 \
    --shiftms 5 \
    --fftl 1024 \
    --mcep_dim_start 2 \
    --mcep_dim_end 27 \
    --mcep_alpha 0.41 \
    --mag 0.5 \
    --inv false \
    --n_jobs 1