techno

Taking a number of techno tracks, we:

• convert to wav, normalize sample rate and tempo
• order tracks using a special correlation measurement
• locally align the clips using a sliding window and finally:
• interpolate to produce a "continuous" mix N.b. still working to improve accuracy of the tempo normalization. Improved tempo normalization would improve xsition quality

example usage:

python3 py/robo_dj.py songs

Notes

• tested on Ubuntu and MacOS
• files in songs/ folder are mp3 or m4a format
• two songs from the larger collection are included in the repo, for testing
• wav conversion, samplerate normalization, tempo normalization, and local alignment steps are performed with local parallelism to take advantage of multicore processors
• Some steps don't re-run if intermediary files already exist: may need to delete intermediary files if you want to reiterate after changing the input data

result

Five hours of high-energy music for long coding, driving, sports or study sessions: https://soundcloud.com/ashrichardson/data-science-mix

2023 Version

https://music.youtube.com/watch?v=FXsyvHnT94U&list=MLPT

findings

• Correlation can be used to align techno clips of same tempo for a smooth transition
• Correlation does not take bar lines or counts into account
• Can get wrong answers like being half a beat off!
• More complex method required for reliability

matrix of correlation between the tracks

clips aligned for correlation on 3.78s window

files:

• robo_dj.py: main subroutine / entry point
• correlation.py: calculate correlations between tracks
• parfor.py: parallel for loop that runs a function in parallel on an array of inputs
• stitch.py: use above correlations to establish ordering, then locally align transitions and interpolate
• work_queue.py: process a list of system calls in parallel; one worker thread per CPU thread

sample input files included:

• first track
• second track

sample output files generated:

• output: continuous mix
• output: transition segment