This repo contains the scripts and demo songs for paper titled "Mixing-Specific Data Augmentation Techniques for Improved Blind Violin/Piano Source Separation". | Paper (arXiv) | Demo Page | Pretrained Models (GoogleDrive) |
The main purpose of this work is to allow users to apply mixing-specific data augmentation techniques to facilitate the training of a neural network model for source separation, in particular with the Open-Unmix model architecture . The training scripts required for training your own models can be found from the official repository of Open-Unmix.
Modification for Training
Note that for our experiment scenario, we modified the model as follows:
- change nb_channels to 1: for monaural cases
- nb-train-samples corresponds to the N in the paper
The contents are oganized as follows:
Pretraind_Models: (Uploaded to GoogleDrive)
- demo page
- 15-sec version of the 16 test songs
- ExtractMedleyDB_16TestSongs.py: for extracting the 16 MedleyDB[2, 3] songs.
Augmentations: demo codes for proposed augmentation methods
- implementation of non-silence is simple, as shown in doc of librosa
Inference.py: demo code for using Pretrained_Models to separate the Demo_mp3_15sec
- SeparateModules.py: modules modified from Open-Unmix test.py
- the required modules should be downloaded from Open-Unmix official repository:
 F.-R. St ̈oter, S. Uhlich, A. Liutkus, and Y. Mitsufuji, “Open-Unmix - Areference implementation for music source separation,”Journal of OpenSource Software, vol. 4, no. 41, p. 1667, 2019.
 MedleyDB Original: R. Bittner, J. Salamon, M. Tierney, M. Mauch, C. Cannam and J. P. Bello, "MedleyDB: A Multitrack Dataset for Annotation-Intensive MIR Research", in 15th International Society for Music Information Retrieval Conference, Taipei, Taiwan, Oct. 2014.
 MedleyDB 2.0 Bittner, R., Wilkins, J., Yip, H., & Bello, J. (2016). MedleyDB 2.0: New Data and a System for Sustainable Data Collection. New York, NY, USA: International Conference on Music Information Retrieval (ISMIR-16).