Note: the code is adapted from r9y9's wavenet vocoder, u can get more information about wavenet at there.
some problems still exists:
- the generated wav from teacher will have some noise in silence area(1000k step)
- the generated wav from student still have little noise, but most high frequence noise have been removed
- use relu rather than leaky relu
- don't apply skip connection after the residual connection, the same as r9y9's implemention
- you should set
share_upsample_conv=Trueinhparams.pywhen u train the student
python preprocess.py \
ljspeech \ # data name, i use ljspeech as defalut
your_data_dir \
the_dir_to_save_data/\
--preset=presets/ljspeech_gaussian.json \
python train.py \
--preset=presets/ljspeech_gaussian.json \
--data-root=your_data_dir \
--hparams='batch_size=9,' \ # in my expreiment, i use 3 gpus(1080Ti)
--checkpoint-dir=checkpoint-ljspeech \
--log-event-path=log-ljspeech
python synthesis.py \
--conditional your_local_condition_path \
--preset=presets/ljspeech_gaussian.json \
your_teacher_checkpoint_path \
your_save_dir
python train_student.py \
--preset=presets/ljspeech_gaussian.json \
--data-root=your_data_dir \
--hparams='batch_size=8,' \ # in my expreiment, i use 4 gpus(1080Ti)
--checkpoint-dir=checkpoint-ljspeech_student \
--log-event-path=log-ljspeech_student \
--checkpoint_teacher=your_teacher_checkpoint_path
python synthesis_student.py \
--conditional your_local_condition_path \
--preset=presets/ljspeech_gaussian.json \
your_checkpoint_path \
your_save_dir
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent