MS500: A NEW VERSION OF MIXING SECRETS DATASET FOR MUSIC SOURCE SEPARATION

This is an accompanying repository of my my master thesis and a submission to Late Breaking Demo of International Society of Music Information Retrieval 2022: MS500: A MULTI-TRACK DATASET FOR HIERARCHICAL MUSIC SOURCE SEPARATION

Citation

@masterthesis{huicheng_zhang_2022_7116102,
  author       = {Huicheng Zhang},
  title        = {{Hierarchical Music Source Separation Using Mixing 
                   Secret Multi-track Dataset}},
  school       = {Universitat Pompeu Fabra},
  year         = 2022,
  month        = sep,
  doi          = {10.5281/zenodo.7116102},
  url          = {https://doi.org/10.5281/zenodo.7116102}
}

The link to my presentation: https://docs.google.com/presentation/d/1BD4iTVwp2obUxl8mThp9g8DmRBhJg3G42zwxMGK_bLg/edit?usp=sharing

Download the published MS500 dataset from zenodo:

The link to the dataset from zenodo is coming soon!! It is big and I need to contact zenodo to increase my quota.

After Downloading the dataset from Zenodo, unzip it and go to my forked slakh-utils repository to convert the audio files from .flac to .wav format (Total ~532GB). Then jump to the Automatic Stem generation and Dataset Formatting section to run ms21_generate_yaml.py

Purpose of this repository:

This repository contains the scripts for constructing MS500. This data is only for research purpose. The material contained in it should not be used for any commercial purpose without the express permission of the copyright holders. Please refer to www.cambridge-mt.com for further details.

For evaluation experiment, our code is in a branch of Asteroid which called ms21.

Constructing MS500 Dataset: Step by Step

• Creat a conda environment:

conda env create -f environment.yml
conda activate ms21db

• Data Gathering from The 'Mixing Secrets' Free Multitrack Download Library;

python geturl.py                               # To generate downloadurls.txt, which will be stored in the same folder
python download_zip.py  "path_to_storage"      # To download the zipped multi-track archives and mp3 mixtures using the urls in downloadurls.txt
python unrar.py "path_to_storage"              # To unzip the zipped archives in the "multitrack_zip" folder and put them in the "unzip_multitrack" folder
python unzip_folder_setup.py "path_to_unzip_multitrack_directory"         # To unify the folder structure

• Semi-automatically annotation generation;

python instrument_classify.py "path_to_unzip_multitrack"                 # To generate annotation .csv

• Basic DSP for audio files :Resampling, zero-padding, "monorizing";

python soundfile_check.py -h
  --path PATH, -p PATH  Dataset directory
  --samplerate SAMPLERATE, -sr SAMPLERATE
                        Target Sampling Rates
  --bitrate BITRATE, -bt BITRATE
                        Target Bit Rate
  --mono, -m            Stereo: False, Mono: True
  --pad, -pd            For zero-padding, boolean

• After splitting the dataset. we can do track-level loudness normalization:

python ms21_normalization.py "path_to_dataset" "output_path" -25      # Default target_loudness = -25 LUFS

• Automatic Stem generation and Dataset Formatting: Same folder structure of MedleyDB and special automatic mixing algorithm for generating stereo backing vocal stem.

python ms21_generate_yaml.py "path_to_the_splitted_and_normalized_dataset" "outpath_to_the_well_structured_dataset" num_threads      #Note that within the script target loudness is set to -25 LUFS

TODO: Design automatic mixing algorithm for each instrument

TODO: A Function to only use non-bleeding multi-tracks for stem generation

TODO: A dataloader for Hierarchical Music Source Separation (HMSS) and MSS training

Statistics of MS500

Instrument Level Bleeding Annotation