A Shepard tone is a stack of sine waves spaced one octave apart, all moving in the same direction, under a loudness curve that does not move. The pitch seems to climb or descend for as long as the sound plays. Roger Shepard described the stepped version in 1964, and Jean-Claude Risset later made the continuous glide that carries both their names.
The mechanism
Draw the spectrum on a logarithmic frequency axis and the partials sit at equal distances, one per octave. A bell-shaped loudness curve covers the whole range: silent at the low edge, loudest at the centre, silent again at the high edge.
Slide every partial upward. The one near the top gets quieter and disappears. A new one enters at the bottom, too quiet to notice. After exactly one octave of travel, each partial stands where its upper neighbour started, so the spectrum is the same as at the beginning.
Because octaves sound like the same note, the ear cannot say which partial is “the” pitch. It follows the direction of motion instead, and that direction never changes. The stage above the controls shows this while the sound plays: the dashed curve stays put and the lines travel under it.
Glide and steps
Glide moves the partials continuously. Speed is the number of seconds one octave takes, which is also the length of one loop. Around 20 to 30 seconds per octave gives the slow, unresolved tension used under film scenes. Hans Zimmer’s score for Dunkirk is the example most often cited. 2 to 4 seconds per octave reads as acceleration, the sound of an engine that keeps revving.
Steps plays the Shepard scale: separate tones, each a semitone from the last. It is a chromatic scale that returns to its first note after 12 steps while seeming to have climbed the whole way. Each step has a 15 ms fade at both ends so the changes do not click.
Centre and octaves
Centre sets where the loudness curve peaks. Octaves sets how wide the curve is, with one partial per octave, so 8 octaves around 500 Hz spans about 31 Hz to 8 kHz.
Wider is more convincing, because partials enter and leave further from the loud middle. The range is kept between 20 Hz and 20 kHz. When the centre and octave count do not fit, the octave count is reduced and the hint under the slider says so.
A loop that closes
One loop is one octave of travel. The file is always a whole number of loops long, and the phase of each partial is computed from its position in the range, not accumulated sample by sample. The last sample of the file leads into the first with the same step as any two neighbouring samples.
That makes one exported loop enough. Set it to repeat in a sampler, a DAW track, or a game audio engine and it runs for as long as the scene needs. Use more loops only when the destination cannot repeat a file.
Keep the loop as WAV. Lossy encoders add a few milliseconds of padding at the start and end of a file, and that padding becomes an audible gap each time the loop comes round.
Speakers that break the illusion
The effect depends on the quiet partials at both ends being reproduced. A phone speaker that drops everything below a few hundred hertz removes the low end of the curve, so new partials appear at an audible level and the reset becomes a small jump once per loop.
For small speakers, raise the centre to 1000 Hz or more and drop to 5 or 6 octaves. On headphones or full-range monitors the default range holds up.
Level
The level slider is a ceiling. The amplitudes under the loudness curve always sum to half the number of partials, wherever they are in their travel, and the output is scaled so that sum equals the chosen level. The signal cannot clip, and measured peaks land a few dB below the setting because the partials rarely all crest together.