Pitch is not a single number. Every DAW represents it as one, MIDI note 69, 440 Hz, +3 semitones, but human pitch perception has two independent dimensions. That distinction matters enough that a Bell Labs researcher named Roger Shepard exploited it in 1964 to create a tone that rises forever, and Hans Zimmer built most of a major film score on the same principle sixty years later.
Two Axes for One Parameter
When you hear a pitch, your auditory system simultaneously extracts two things:
Pitch height is the global position on a continuous low-to-high scale. A 880 Hz tone is higher than a 220 Hz tone in an absolute, measurable sense. Height requires knowing where in the audible spectrum you are.
Pitch chroma, or pitch class, is which note within an octave. C4, C5, and C6 share the same chroma. When you hum a familiar melody transposed up an octave, the height changes completely but the chroma sequence, and with it the emotional contour and identity of the tune, remains intact.
These two dimensions are partially independent, and they are not equally robust. Chroma is extracted with high confidence even when frequency information is partial or masked. Height perception is fragile: it depends on reliable access to a tone's absolute position in the spectrum. When that context is removed, the brain defaults to tracking chroma only, and loses its footing on where it is in pitch space.
That fragility is the exploit.
The Bell Labs Experiment, 1964
Roger Shepard was working at Bell Telephone Laboratories in Murray Hill, New Jersey when he published "Circularity in Judgments of Relative Pitch" in the Journal of the Acoustical Society of America (Vol. 36, No. 12, December 1964). The paper describes an experiment using computer-generated tones, each built from 10 sinusoidal partials spaced exactly one octave apart.
A single tone with 10 octave-separated partials sounds complex and spectrally ambiguous. You hear a chroma clearly, because multiple octaves share the same pitch class, but height is uncertain. Which of the ten partials is "the fundamental" is undefined, because they all carry equal weight as candidates.
Shepard arranged 12 such tones a semitone apart, covering one full chromatic octave. Played in ascending order, each step sounds unambiguously higher than the one before. After 12 steps, you are back to the starting pitch class, but listeners consistently reported hearing a continuous ascent with no moment of falling. Transitivity had broken down: if tone A sounds higher than B, B higher than C, and so on around the 12-step circle, then eventually L sounds higher than K and K sounds higher than A.
The mechanism is the spectral envelope. Each partial's amplitude is shaped by a fixed Gaussian bell curve centered in the audible range and plotted on a log-frequency axis:
A_k = exp( -(log2(f_k / f_center))^2 / (2 * sigma^2) )where f_k is the frequency of partial k, f_center is the target center of the bell, and sigma controls the width in octaves. Shepard used roughly 1.5 octaves for sigma, covering around 3 octaves at half-power on each side of center.
As a tone ascends through the sequence, its highest partials climb past the peak of the Gaussian and begin fading, while new low-frequency partials enter from below and fade in. The spectral center of gravity stays pinned. There is no audible moment where pitch drops: the fading high component and the growing low component trade places invisibly. Height remains perpetually undefined because the loudest energy is always in the same spectral region, regardless of which nominal frequencies are playing.
The Continuous Version: Risset's Glissando
Jean-Claude Risset (1938-2016), a French composer and researcher, arrived at Bell Labs in 1965. Working with the same computational infrastructure Shepard had used, Risset extended the discrete 12-step staircase into a continuous, smoothly sweeping glissando, completed during his time at Bell Labs through 1969.
Where Shepard's tone clicks up semitone by semitone, the Risset glissando sweeps all partials upward as a continuous frequency modulation while the Gaussian amplitude function re-evaluates in real time. As the highest partial becomes inaudible, the lowest partial has already faded in below. When looped, the descent or ascent runs indefinitely with no audible seam.
The synthesis requirement is minimal: a set of oscillators locked in octave ratios, each amplitude-controlled by the Gaussian function evaluated against its instantaneous log-frequency position. The function runs continuously. There is no moment of crossfade logic, no special transition handling. The math handles the handoff automatically.
Three Simultaneous Scales: Dunkirk
Christopher Nolan asked Hans Zimmer to write a Dunkirk (2017) score that felt like it was "perpetually building." Zimmer anchored nearly the entire film to a Shepard tone derived from the film's central sound design motif: a ticking watch.
The approach is more structured than simply looping a glissando. Zimmer used three simultaneous Shepard scales running at different rates, each representing one of the film's three interwoven timescales. The slow rate corresponds to the week the soldiers waited on the beach. The medium rate corresponds to the day the civilian boats were at sea. The fast rate corresponds to the hour the Spitfire pilots were airborne. Tension accumulates within each layer independently and multiplicatively, as the different rates phase in and out of alignment, creating momentary convergences that the score exploits for dramatic peaks.
The technique transfers directly to game audio and interactive scoring, where tension must scale non-linearly with player state. Multiple Shepard rates can be layered and mixed in proportion to threat level, timer pressure, or proximity to a goal, producing continuously building intensity without the artifact of a looping piece that sounds like it restarts.
In The Dark Knight (2008), sound designer Richard King used the effect differently. The Batpod's engine acceleration sound was constructed from a Shepard tone played on a keyboard. As King described it: "When played on a keyboard, it gives the illusion of greater and greater speed; the pod appears unstoppable." The vehicle's pitch never peaks because the spectral ceiling is hidden. It simply accelerates.
The Tritone Paradox
Diana Deutsch, a music psychologist at UC San Diego, reported a related effect in 1986 that extends Shepard's framework in an unexpected direction. When listeners hear a tritone interval, two Shepard tones separated by exactly half an octave, the interval is spectrally ambiguous about direction. Because the two tones are equidistant on the pitch-class circle, there is no harmonic cue to distinguish ascending from descending.
Deutsch found that listeners' judgments split based on their native language and regional dialect. English speakers from California tended to hear specific tritone pairs as ascending; British English speakers from southern England tended to hear the same pairs as descending. The correlation holds at the group level and replicates across studies.
This is not a minor calibration difference. It reveals that the internal spectral templates used to resolve pitch height ambiguity, the prior that fills in when chroma is clear but height is not, are shaped by the tonal statistics of the listener's native speech. The language you grew up speaking has quietly tuned the same system Shepard exploited, and a tone that sounds "obviously rising" to one engineer may sound "obviously falling" to the person sitting next to them.
Building One in Any DAW
A functional Shepard tone requires no specialist tools:
- Stack 8 to 10 oscillators tuned to successive octaves across the audible range. A typical spread: 27.5, 55, 110, 220, 440, 880, 1760, 3520, 7040 Hz.
- Apply a single pitch LFO to all oscillators simultaneously, with identical depth and rate. Their octave spacing must remain constant as they sweep.
- Set each oscillator's output gain using the Gaussian formula above, evaluated against that oscillator's current frequency. Fixed gain values work for a discrete scale; automated gain tracks the sweep continuously for a glissando.
- As the LFO cycles, the highest oscillators periodically fold back to the bottom of the range. Their Gaussian amplitude at re-entry is near zero, so the fold is inaudible.
In Max/MSP the standard implementation uses a single phasor driving all oscillator pitches through [mtof] and octave offsets, with a [cos~] computing the Gaussian amplitude coefficient from each oscillator's log-frequency distance to a fixed center. In Ableton or Bitwig, a Utility device on each of 8 audio channels, volume-automated by an LFO that mirrors the Gaussian position, achieves the same result without code.
The rate matters. Below roughly one semitone per second, listeners have time to track absolute pitch and the illusion weakens. Above three to four semitones per second, the ascent becomes disorienting and overtly artificial. The Zimmer Dunkirk rates, tuned to the three timescales of the film, are slow enough to work subliminally across the full running time.
What It Reveals
Shepard's experiment did not just produce an interesting demo. It proved that pitch height and pitch chroma are processed by separable mechanisms, that one can be rendered unreliable while the other remains sharp, and that the result is a systematic perceptual failure. The brain is not detecting a trick and failing to correct it. It genuinely does not have access to the information that would reveal the loop.
The same architecture is present in every pitch judgment you make. Height perception is always partially inferred from context. Shepard tones work because they remove the context, not because they deceive a system that has it.