Which version do you prefer?

· algorithm, Music

I recently worked on the alignment of sequences, using the dynamic time warping (DTW). Cleaning up my drive, I found two versions of the French song “Comme un Boomerang“, written and composed by Serge Gainsbourg. One version is the well known duo version between Dani (the original recipient of the song) and Etienne Dao, and the other version is a recently discovered version by Gainsbourg himself.

Both versions are very similar, in style, with almost the same beat. Just out of interest, I checked whether I could use my DTW algorithm to align these two tracks. Although the algorithm was designed for speech signals, the result is quite convincing. It is also (slightly) better than aligning the tracks manually (actually, the tempos seem to fluctuate a bit, which makes that alignment a bit tedious…).

Here is the resulting alignment: left channel is the Gainsbourg version, and right channel is the Dani/Dao one. Choose the one you prefer! Or just enjoy it as a trio version of the tune…


Technical note: The alignment is essentially based on MFCCs and energy. The tracks are realigned and synthesized through a method that takes its roots in Phase Vocoder theory, coupled with the overlap-and-add technique. There are quite a few artifacts due to this synthesis method. Anyone with a better idea is welcome to apply it…

Edit (19/6/2012): Modified the resynthesis procedure, using something closer to the phase vocoder by Ellis (available here). Thanks to Cyril Joder for bringing that up! Below is the resulting alignment (well, same alignment, different rendering):

Subjectively, it feels more natural on the right channel (Dao/Dani version) while the left channel is “perfect”: apart from being the Gainsbourg version (of course), the left channel is the original audio since I did not modify its tempo. Basically this reflects the fact that the resynthesis guarantees a perfect reconstruction – was not the case with the instantaneous frequency version, above.

Leave a Comment