Stretch a recording until its notes become weather. Paulstretch reads the sound in long, overlapping windows, throws away the timing inside each one, and writes it back out many times slower, so a song turns into a slowly shifting drone in the same key.
How Paulstretch works
The algorithm is Nasca Octavian Paul’s. It reads the input one window at a time, a few hundred milliseconds long, takes the window’s spectrum, and keeps only the magnitudes. Every frequency gets a new random phase, and the window is turned back into sound.
The trick is in how fast the two sides move. The output advances half a window per step; the input advances half a window divided by the stretch. At 8x, each output step moves the input along by 1/16 of a window, so neighbouring windows hear almost the same audio and blend into a continuous tone.
Because every phase is new, nothing sharp survives. A drum hit is a burst of phases lined up at one instant, and once the phases are random it is a brief rise in noise. That is why Paulstretch sounds smooth where other stretchers sound metallic, and why it is useless for keeping a beat.
Stretch amount
| Stretch | What happens to a song |
|---|---|
| 2 to 4x | Still recognisable, words slurred, rhythm gone |
| 8x | The classic setting: chords bloom and hang |
| 20 to 50x | One chord lasts minutes; the song becomes a texture |
The output length is shown under the settings. Stretch a short range for a usable result: a ten-second chorus at 30x is five minutes of pad.
Window size
The window is rounded to a power of two samples, which the FFT needs, so the menu shows the real length at the file’s sample rate. Short windows keep more of how the sound moves but add a grainy, phasey texture. Long windows average more of the music into each step: a chord change becomes a slow crossfade and voices lose their consonants entirely.
Level and stereo
Successive windows share no phase, so they add as power, not as amplitude. The windows are shaped so their power sums to a constant, and the output sits at about the input’s average level without any normalising pass. Peaks are lower than the source’s, because a drone has no transients, so the result usually sounds quieter at the same RMS.
Keep the stereo image gives both channels the same random phases, which keeps anything panned in place. Turned off, each channel is randomised separately and the result spreads wide, which suits ambient work and loses level in mono.
Long files
The export decodes the source once, front to back, and writes each finished window as it goes, so memory stays at a few windows whatever the length. What does grow is the output: plan for it, since a stretched album can run to many hours and gigabytes.