Cutting a 90-minute DJ set down to 60 without killing the transitions

A friend asked me to turn a 92-minute recorded set into a 60-minute version for a radio slot. The source was a single WAV, 48 kHz/24-bit, about 1.2 GB, with no cue sheet and no tracklist — just a file and a photo of the running order scrawled on the back of a receipt. The mix was mostly continuous, so there were no clean gaps to cut at, only transitions and a few breakdowns.

I did not try to cut on the beat. Instead I marked every transition in Audacity by looking at the spectral view and listening at 1.5x, then exported each candidate cut point as a short reference clip. For the actual edits I used ffmpeg with a filter that trims at the sample level and crossfades the seam, because a hard splice at a transition is audible even when the tempo matches.

ffmpeg -i set.wav -filter_complex "\
[0:a]atrim=0:1832.4,asetpts=PTS-STARTPTS[a];\
[0:a]atrim=1834.9:5520.0,asetpts=PTS-STARTPTS[b];\
[a][b]acrossfade=d=0.35:c1=tri:c2=tri[out]" \
-map "[out]" -c:a pcm_s24le set_60.wav

That 0.35 s crossfade is the whole trick. Long enough to hide the phase jump when two tracks overlap, short enough that the incoming track’s downbeat still lands where the crowd expects it. When I tried 0.8 s the kick doubled and it sounded like a flam; at 0.1 s there was a click on every cut.

The cost is that you lose the tail of whatever was playing. On a mix where the outgoing track is doing a long filter sweep, cutting 2.5 s early is noticeable if you know the original. I kept a spreadsheet of each cut with the timestamp and what got lost, and for two of the eleven cuts I moved the seam later into the breakdown instead, where the loss is masked by a riser.

Where this falls apart: tracks with live vocals or unquantized drums. Crossfading across a sung phrase sounds like a stutter, and a human drummer drifts enough that the 0.35 s overlap beats against itself. I had one funk section where I ended up cutting whole bars at a phrase boundary instead of crossfading — slower, but the only version that didn’t sound edited. Also, anything with a sustained bass note across the seam will phase-cancel; check the low end in a spectrum analyzer before committing.

The 60-minute version went out and nobody mentioned the edits, which I took as the point. But I still wonder whether the two cuts I moved into breakdowns are actually the weaker ones — they’re the ones I notice on a relisten. How do you decide between a seam that’s technically cleaner and one that preserves the phrase you’d want to hear?