crank

Non-parallel voice conversion based on vector-quantized variational autoencoder with adversarial learning

Setup

• Install Python dependency

$ git clone https://github.com/k2kobayashi/crank.git
$ cd crank/tools
$ make

• install dependency for mosnet

$ sudo apt install ffmpeg   # mosnet dependency

Recipes

• English
  • VCC2020
  • VCC2018 (Thanks to @unilight)
• Japanese
  • jsv_ver1

Conversion samples

You can access several converted audio samples of VCC 2018 dataset in the URL.

• vcc2020v1
• vcc2018v1

Run VCC2020 recipe

crank has prepared recipe for Voice Conversion Challenge 2020. In crank recipe, there are 7 stages to implement non-parallel voice conversion.

• stage 0
  • download dataset
• stage 1
  • initialization
    • generate scp files and figures to be determine speaker-dependent parameters
• stage 2
  • feature extraction
    • extract mlfb and mcep features
• stage 3
  • training
• stage 4
  • reconstuction
    • generate reconstructed feature for fine-tuning of neural vocoder
• stage 5
  • evaluation
    • convert evaluation waveform
• stage 6
  • synthesis
    • synthesis waveform by pre-trained ParallelWaveGAN
    • synthesis waveform by GriffinLim
• stage 7
  • objective evalution
    • mel-cepstrum distortion
    • mosnet

Put dataset to downloads

Note that dataset is only released for the participants (2020/05/26).

$ cd egs/vaevc/vcc2020v1
$ mkdir downloads && cd downloads
$ mv <path_to_zip>/vcc2020_{training,evaluation}.zip downloads
$ unzip vcc2020_training.zip
$ unzip vcc2020_evaluation.zip

Run feature extraction and model training

Because the challenge defines its training and evaluation set, we have initially put configuration files. So, you need to run from 2nd stage.

$ ./run.sh --n_jobs 10 --stage 2 --stop_stage 5

where the n_jobs indicates the number of CPU cores used in the training.

Configuration

Configurations are defined in conf/mlfb_vqvae.yml. Followings are explanation of representative parameters.

• feature

When you create your own recipe, be carefull to set parameters for feature extraction such as fs, fftl, hop_size, framems, shiftms, and mcep_alpha. These parameters depend on sampling frequency.

• feat_type

You can choose feat_type either mlfb or mcep. If you choose mlfb, the converted waveforms are generated by either GllifinLim vocoder or ParallelWaveGAN vocoder. If you choose mcep, the converted waveforms are generated by world vocoder (i.e., excitation generation and MLSA filtering).

• trainer_type

We support training with vqvae, lsgan, cyclegan, and stargan using same generator network.

• vqvae: default vqvae setting
• lsgan: vqvae with adversarial learning
• cyclegan: vqvae with adevesarial learning and cyclic constraints
• stargan: vqvae with adevesarial learning similar to cyclegan

Create your recipe

Copy recipe template

Please copy template directory to start creation of your recipe.

$ cp -r egs/vaevc/template egs/vaevc/<new_recipe>
$ cd egs/vaevc/<new_recipe>

Put .wav files

You need to put wav files appropriate directory. You can choose either modifying download.sh or putting wav files. In either case, the wav files should be located in each speaker like following <new_recipe>/downloads/wav/{spkr1, spkr2, ..., spkr3}/*.wav.

If you modify downaload.sh,

$ vim local/download.sh

If you put wav files,

$ mkdir downloads
$ mv <path_to_your_wav_directory> downloads/wav
$ touch downloads/.done

Run initialization

The initialization process generates kaldi-like scp files.

$ ./run.sh --stage 0 --stop_stage 1

Then you modify speaker-dependent parameters in conf/spkr.yml using generated figures. Page 20~22 in slide help you how to set these parameters.

Run feature extraction, train, reconstruction, and evaluation

After preparing configuration, you run it.

$ ./run.sh --stage 2 --stop_stage 7

Citation

Please cite this paper when you use crank.

K. Kobayashi, W-C. Huang, Y-C. Wu, P.L. Tobing, T. Hayashi, T. Toda,
"crank: an open-source software for nonparallel voice conversion based on vector-quantized variational autoencoder",
Proc. ICASSP, 2021. (accepted)

Achknowledgements

Thank you @kan-bayashi for lots of contributions and encouragement helps.

Who we are

• Kazuhiro Kobayashi @k2kobayashi [maintainer, design and development]

• Wen-Chin Huang @unilight [maintainer, design and development]

• Tomoki Toda [advisor]