Can you speak like a violin ?

How to install dependencies ?

pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html

---

How to load the dataset ?

Use src\timbre_transfer\datasets\NSynthDataset\NSynthDataset.py :

Example :

from src.timbre_transfer.datasets.NSynthDataset import NSynthDataset
root_dir = 'data'
train_dataset = NSynthDataset(root_dir = root_dir, usage = 'train', select_class='vocal_acoustic', transform=None)

Dataset files hierarchy (here, the root_dir is named "data", as shown on the Python example above) :

• data/
  • nsynth/
    • nsynth-test/
      • examples.json
      • audio/
        • All audio files (.wav) used for testing
    • nsynth-train/
      • examples.json
      • audio/
        • All audio files (.wav) used for training
    • nsynth-valid/
      • examples.json
      • audio/
        • All audio files (.wav) used for validating

---

source /miniconda/bin/activate

---

How to train the model ?

2VAEs : python train_2VAEs.py

2VAEs + CC : python train_2VAEs_CC.py

2VAEs + CC + GAN : python train_2VAEs_CC_GAN.py

---

How to export some sounds and spectrograms ?

2VAEs : python exports_2VAE.py

2VAEs + CC : python exports_2VAE_CC.py

2VAEs + CC + GAN : python exports_2VAE_CC_GAN.py

---

Audio results presented in https://gbittencourt.github.io/ATIAM2022-timbre-transfer/

---