chainer-ClariNet

A Chainer implementation of ClariNet( https://arxiv.org/abs/1807.07281 ).

Results

Autoregressive WaveNet(Single Gaussian ver.) trained with VCTK Corpus

https://nana-music.com/sounds/04027269/

Student Gaussian IAF trained with LJ-Speech

https://nana-music.com/sounds/043ba7b4/

Requirements

I trained and generated with

• python(3.5.2)

• chainer (5.0.0b4)

• librosa (0.6.2)

• matplotlib (2.2.3)

• tqdm (4.25.0)

Usage

download dataset

You can download VCTK Corpus(en multi speaker)/LJ-Speech(en single speaker) very easily via my repository.

set parameters

Almost parameters in params.py and teacher_params.py are same as params.py in my other repositories like VQ-VAE. If you modified params.py in AutoregressiveWavenet, you have to replace teacher_params.py with it to train student.

training

You can use same command in each directory.

(without GPU)
python train.py

(with GPU #n)
python train.py -g n

You can resume snapshot and restart training like below.(Now support AutoregressiveWaveNet only)

python train.py -r snapshot_iter_100000

Other arguments -f and -p are parameters for multiprocess in preprocessing. -f means the number of prefetch and -p means the number of processes. I highly recommend to modify -f to large number like 64. If GPU-Util is stil low, modify -p to large number like 8.

generating

python generate.py -i <input file> -o <output file> -m <trained model>

If you don't set -o, default file name result.wav is used. If you don't set -s, the speaker is same as input file that got from filepath.

Caution

I only check the results for

• Autoregressive WaveNet(Single Gaussian ver.)

• Student Gaussian IAF