Why does stem separation take minutes rather than seconds?

Product support · SubverseLab

The model reads the track in fixed 7.8-second windows and those windows overlap by a quarter, so a four-minute track is about 41 passes rather than 31. The overlap is not decoration: without it the windows are butt-joined and every seam leaves a click — measured at 35 times the surrounding signal on real music, worst in the vocal. With the cross-fade it drops to 2.4, which is indistinguishable from ordinary signal. So a separation that finishes quickly is one that left the ticks in.

Answers

• If you only need the tempo and the key, use Key & BPM instead — that reads the whole file in well under a second and is always free.