DiffSinger / docs /README-SVS-popcs.md
ddd
Add application file
b93970c
|
raw
history blame
3.57 kB

DiffSinger (SVS version)

0. Data Acquirement

1. Preparation

Data Preparation

a) Download and extract PopCS, then create a link to the dataset folder: ln -s /xxx/popcs/ data/processed/popcs

b) Run the following scripts to pack the dataset for training/inference.

export PYTHONPATH=.
CUDA_VISIBLE_DEVICES=0 python data_gen/tts/bin/binarize.py --config usr/configs/popcs_ds_beta6.yaml
# `data/binary/popcs-pmf0` will be generated.

Vocoder Preparation

We provide the pre-trained model of HifiGAN-Singing which is specially designed for SVS with NSF mechanism. Please unzip this file into checkpoints before training your acoustic model.

(Update: You can also move a ckpt with more training steps into this vocoder directory)

This singing vocoder is trained on ~70 hours singing data, which can be viewed as a universal vocoder.

2. Training Example

First, you need a pre-trained FFT-Singer checkpoint. You can use the pre-trained model, or train FFT-Singer from scratch, run:

# First, train fft-singer;
CUDA_VISIBLE_DEVICES=0 python tasks/run.py --config usr/configs/popcs_fs2.yaml --exp_name popcs_fs2_pmf0_1230 --reset
# Then, infer fft-singer;
CUDA_VISIBLE_DEVICES=0 python tasks/run.py --config usr/configs/popcs_fs2.yaml --exp_name popcs_fs2_pmf0_1230 --reset --infer 

Then, to train DiffSinger, run:

CUDA_VISIBLE_DEVICES=0 python tasks/run.py --config usr/configs/popcs_ds_beta6_offline.yaml --exp_name popcs_ds_beta6_offline_pmf0_1230 --reset

Remember to adjust the "fs2_ckpt" parameter in usr/configs/popcs_ds_beta6_offline.yaml to fit your path.

3. Inference Example

CUDA_VISIBLE_DEVICES=0 python tasks/run.py --config usr/configs/popcs_ds_beta6_offline.yaml --exp_name popcs_ds_beta6_offline_pmf0_1230 --reset --infer

We also provide:

  • the pre-trained model of DiffSinger;
  • the pre-trained model of FFT-Singer for the shallow diffusion mechanism in DiffSinger;

Remember to put the pre-trained models in checkpoints directory.

Note that:

  • the original PWG version vocoder in the paper we used has been put into commercial use, so we provide this HifiGAN version vocoder as a substitute.
  • we assume the ground-truth F0 to be given as the pitch information following [1][2][3]. If you want to conduct experiments on MIDI data, you need an external F0 predictor (like MIDI-old-version) or a joint prediction with spectrograms(like MIDI-new-version).

[1] Adversarially trained multi-singer sequence-to-sequence singing synthesizer. Interspeech 2020.

[2] SEQUENCE-TO-SEQUENCE SINGING SYNTHESIS USING THE FEED-FORWARD TRANSFORMER. ICASSP 2020.

[3] DeepSinger : Singing Voice Synthesis with Data Mined From the Web. KDD 2020.