Taejun Kim, Yi-Hsuan Yang and Juhan Nam
"Reverse-Engineering The Transition Regions of Real-World DJ Mixes using Sub-band Analysis with Convex Optimization"
Proceedings of the New Interfaces for Musical Expression (NIME), 2021 [Slides]

Code for WEB DEMO can be found here.

In this example, we are going to use the awesome mix below by Palms Trax since it is one of the author's favorite mixes. data/meta/tracklist.csv contains the tracklist of the mix which is manually collected by the author from YouTube comments.

NOTE: This study and repository are extensions of our previous study and repository, mix-to-track subsequence alignment, a preprocessing step of this transition analysis.

You can get visualizations of transition analysis as shown below after running scripts:

and audio files reproduced by the algorithm under results directory:

# An example directory of result outputs of the first transition
results/00000-cPo-qzbGLqE-05-RocnNoPCZDw-urJv3U_fgCY
├── mix_djmixed.mp3  # transition by the DJ (Palms Trax)
├── mix_reproduced-base.mp3  # transition by baseline
├── mix_reproduced-eq3-subscale.mp3  # transition by our method (EQ-3) 
├── mix_reproduced-xfade.mp3  # transition by previous method (crossfade)
├── next_reproduced-base.mp3  # baseline only applied to the next track
├── next_reproduced-eq3-subscale.mp3  # our method only applied to the next track
├── next_reproduced-xfade.mp3  # previous method only applied to the next track
├── prev_reproduced-base.mp3  # baseline only applied to the previous track
├── prev_reproduced-eq3-subscale.mp3  # our method only applied to the previous track
├── prev_reproduced-xfade.mp3  # previous method only applied to the previous track
├── viz-base.pdf  # visualization of baseline
├── viz-eq3-subscale.pdf  # visualization of our method
└── viz-xfade.pdf  # visualization of previous method

NOTE: This repo is written and tested on Python 3.8.10.

You can install required Python packages using the code below:

pip install -r requirements.txt
pip install madmom==0.16.1  # madmom should be installed after installing cython

The scripts should be run in order as described below:

1. python scripts/download.py downloads audio files from YouTube.
2. python scripts/feature_extraction.py extracts features and saves them under cache/ using disk-caching of joblib.
3. python scripts/alignment.py performs mix-to-track subsequence DTW and saves the alignment results.
4. python scripts/segmentation.py performs mix segmentation and saves the results.
5. python scripts/reproduce_mixing.py finally performs transition analysis using convex optimization.