Record the input in grains of a fixed length and play each grain back reversed, one grain behind the source. Every phrase gains a rising copy of itself, which is the pre-echo swell that used to require flipping a tape reel and playing it the wrong way round.
Grain length
Grain sets how much audio is reversed at a time, from 50 ms to 2000 ms. It decides two things at once: how long each swell lasts, and how far behind the source it arrives.
Short grains, under 150 ms, read as a stutter or a smear rather than a swell. Around 300 ms to 600 ms the effect is recognisably backwards. Past 1 second the reversed copy drifts far enough behind the source to work as a separate part.
The grain is also the unit of timing. At 120 BPM a beat is 500 ms, so a 500 ms grain puts one swell per beat and a 375 ms grain puts one per dotted eighth. The delay calculator converts a tempo to milliseconds if you need a division other than the obvious ones.
Feedback and how it alternates
Feedback writes the reversed output back into the next grain instead of only the dry input. Reversing something a second time turns it forwards again, so repeats alternate direction: the first repeat is backwards, the second is forwards but shifted, the third is backwards again.
At low settings this reads as a longer swell. Above about 60% it turns into a loop that keeps re-photographing itself, which is characteristic and also the point at which the effect stops being subtle.
Stereo spread
Spread lengthens the right channel’s grain by up to a quarter, so the two sides run at slightly different rates and slide in and out of alignment instead of swelling together. At 0 both channels use the same grain and the effect stays centred.
The length is only ever changed at a grain boundary, where the fade at the seam is already at zero. Moving the control during playback does not tear the grain in flight.
Seams
Each grain is faded in and out over 25 ms. Without that fade, a reversed grain would start on whatever sample happened to end the previous one, and every boundary would be a click.
The fade is also why very short grains lose level: at 50 ms, the two fades take up most of the grain. If short settings sound quieter than expected, this is why.
Mix
Mix balances the dry signal against the reversed copy. At 0 the audio passes through untouched. Above 70% the source starts to disappear behind its own swells, which suits a send more than a whole track.