Implementation of the DDSP model using PyTorch. This implementation can be exported to a torchscript model, ready to be used inside a realtime environment (see this video).

This implementation is specially for generating cat voices!

Edit the config.yaml file to fit your needs (audio location, preprocess folder, sampling rate, model parameters...), then preprocess your data using

python preprocess.py

You can then train your model using

python train.py --name mytraining --steps 10000000 --batch 16 --lr .001

Once trained, export it using

python export.py --run runs/mytraining/

It will produce a file named ddsp_pretrained_mytraining.ts, that you can use inside a python environment like that

import torch

model = torch.jit.load("ddsp_pretrained_mytraining.ts")

pitch = torch.randn(1, 200, 1)
loudness = torch.randn(1, 200, 1)

audio = model(pitch, loudness)

cat voice segmentation: (Please change paths and parameters inside the code)

python seg_cat.py

Pitch shifting:

git clone https://github.com/ederwander/PyAutoTune.git
cd PyAutoTune
pip install -e .
cd ../ddsp_pytorch
python TuneAndSaveToFile_folder.py data/cats/train/

file inference:

python inference_file.py --model_path export/ddsp_cats_seg_pitchshift_pretrained.ts --input_file violin_original.mp3

folder inference (for evaluation only, please change paths and parameters inside the code):

python inference_folder.py

python evaluate_f0.py [output folder] [reference folder]