Timbre transfer samples can be found here.

1. Install anaconda/miniconda
2. Run in anaconda prompt: conda env create -f environment.yml
3. Activate new environment: conda activate timbrer

Run generate_set_singlethread.py. The input midi files will be downloaded and a dataset of mel spectrogram pairs will be created. Copy the resulting files from data/spectrogram to pix2pixHD/datasets/timbre.

Modify pix2pixHD/train.py by setting the correct test and dataset names. Start the training process: cd pix2pixHD/ && python train.py. Intermediate results will be visible in the pix2pixHD/checkpoints/<test_name>/web/folder.

Modify pix2pixHD/test.py by setting the test and dataset names. Run the file, which will generate the output spectrograms into the pix2pixHD/results/<test_name> folder.

Run reconstruct.py and give it a list of .npy spectrograms or directories containing spectrograms. Optimizes white noise using L-BFGS to produce the target spectrogram. Expect around ~2s/image on a modern GPU. The resulting waveforms will be placed in the same folder as the inputs.

This work is a derivative of pix2pixHD by NVIDIA Corporation.