HiFiPLN

Multispeaker Community Vocoder model for DiffSinger

This is the code used to train the "HiFiPLN" vocoder.

A trained model for use with OpenUtau is available for download on the official release page.

Why HiFiPLN?

Because a lot of PLN was spent training this thing.

Training

Python

Python 3.10 or 3.11 is required.

Data preparation

Preperocessing and splitting the dataset into smaller files is done using a single script. Note that if the input files are shorter than --length seconds, they will be skipped. It is better to provide full unsegmented files to the script, but if your input files are already split into chunks, you can run with --length 0 to disable splitting.

python preproc.py --config PATH_TO_CONFIG -o "dataset/train" --length 1 PATH_TO_TRAIN_DATASET

You will also need to provide some validation audio files. Run preproc.py with --length 0 to disable segmenting.

python preproc.py --config PATH_TO_CONFIG -o "dataset/valid" --length 0 PATH_TO_VALIDATION_DATASET

Train model

python train.py --config "configs/hifipln.yaml"

• If you see an error saying "Total length of `Data Loader` across ranks is zero" then you do not have enough validation files.
• You may want to edit configs/hifipln.yaml and change train: batch_size: 12 to a value that better fits your available VRAM.

Resume

python train.py --config "configs/hifipln.yaml" --resume CKPT_PATH

You may set CKPT_PATH to a log directory (eg. logs/HiFiPLN), and it will find the last checkpoint of the last run.

Finetuning

Download a checkpoint from https://utau.pl/hifipln/#checkpoints-for-finetuning \ Save the checkpoint as ckpt/HiFiPLN.ckpt then run:

python train.py --config "configs/hifipln-finetune.yaml"

• Finetuning shouldn't be run for too long, especially for small datasets. Just 2-3 epochs or ~20000 steps should be fine.

Exporting for use in OpenUtau

python export.py --config configs/hifipln.yaml --output out/hifipln --model CKPT_PATH

You may set CKPT_PATH to a log directory (eg. logs/HiFiPLN), and it will find the last checkpoint of the last run.

Credits

• DiffSinger
• Fish Diffusion
• HiFi-GAN (Paper)
• OpenVPI
• PC-DDSP
• RefineGAN
• UnivNet (Paper)
• FA-GAN
• APNet2