Here, you will find the code related to PU-Iter algorithm for source separation.
If you use any of the things existing in this repository, please cite the corresponding paper.
You can also find an online demo with sound examples related to this work on the companion website.
To reproduce the experiments conducted in our paper, you will need to download the Dexmixing Secret Database (DSD100) and to place its content in the dataset/DSD100 folder.
If you use this dataset, you will end up with the proper directory structure and file names, as used in the functions/get_data_DSD function.
If you want to use a different dataset, then you have two options:
- either you format your file names and directory structure to match the one from DSD100;
- or you modify the file reading function
get_data_DSDto suit your needs.
We also provide piano notes from the MAPS database, as well as the corresponding loading function get_data_MAPS_notes. This is intended to reproduce the results of the article on piano sounds, which corresponds to the script plot_res_piano.
The script to reproduce the experiments are placed in the scripts folder. They will notably record audio files in the audio_files folder, and some metrics (SDR, SIR and SAR, but also inconsistency or various cost functions) in the metrics folder.