Granular Synthesis

In 1947, the physicist Dennis Gabor proposed that any sound can be decomposed into a cloud of elementary acoustic particles — short bursts he called quanta — each carrying a tiny slice of frequency and time. The idea was theoretical for decades. Then digital computing caught up, and granular synthesis was born.

The core idea is disarmingly simple: take any sound, chop it into grains of 1–100 milliseconds, and scatter those grains back out in time. Change their density, pitch, position, and overlap, and the resulting texture can be a whisper, a drone, a stretched-out vowel frozen in amber, or something that has no name in any instrument family.

Because each grain is so short, the human ear cannot hear it as a discrete event — it perceives the statistical cloud instead. That gap between what the algorithm does (schedule individual grains) and what you hear (a continuous texture) is where granular synthesis hides its computational depth.

Each dot below is a grain — a brief burst of sound lasting just a few milliseconds. Adjust the controls to change how many grains fire per second, how wide their pitch is spread, and how long each one lasts.

<!-- {{c_intro}} -->
<p class="hint">{{hint_para}}</p>
<canvas id="canvas" width="560" height="240" aria-label="{{canvas_label}}"></canvas>
<div class="controls">
  <label>{{lbl_density}}<input type="range" id="density" min="1" max="60" value="20"></label>
  <label>{{lbl_pitch}}<input type="range" id="pitch" min="0" max="100" value="30"></label>
  <label>{{lbl_duration}}<input type="range" id="duration" min="5" max="120" value="40"></label>
</div>
<div class="status" id="status"></div>
<div class="btns">
  <button id="toggle" type="button">{{btn_start}}</button>
  <button id="reset" type="button" class="ghost">{{btn_reset}}</button>
</div>

/* {{c_layout}} */
* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; color: #222; margin: 0; }
.hint { font-size: .9rem; color: #444; margin: 0 0 .7rem; line-height: 1.45; }
canvas { display: block; background: #0d1117; border-radius: 8px; width: 100%; max-width: 560px; margin: 0 auto .8rem; }
.controls { display: flex; flex-direction: column; gap: .4rem; margin-bottom: .6rem; font-size: .88rem; color: #333; }
label { display: flex; align-items: center; gap: .6rem; }
input[type=range] { flex: 1; accent-color: #1d3557; }
.status { font-size: .95rem; font-weight: 600; min-height: 1.4em; margin: .3rem 0; color: #1d6a3a; }
.btns { display: flex; gap: .5rem; flex-wrap: wrap; }
button { font: 600 14px system-ui, sans-serif; padding: .45rem .9rem; border: 1px solid #1d3557;
         background: #1d3557; color: #fff; border-radius: 8px; cursor: pointer; }
button.ghost { background: #fff; color: #1d3557; }

// Code not found

Notice what happens at extremes. With very few grains per second you can hear the individual bursts — the rhythm of scheduling becomes audible. Pack them densely enough and the ear fuses them into a seamless cloud. Density, pitch scatter, and grain duration are the three knobs that turn raw milliseconds into any texture from airy shimmer to grinding noise.

Playing back thousands of grains per second in real time is not as simple as pressing play on each one.

• Scheduling: every grain has a start time, duration, pitch, amplitude envelope, and panning position. A synthesizer must compute and mix hundreds of grains simultaneously — each one a short convolution of an envelope and a waveform window (typically a Hann or Gaussian bell).
• Overlap and polyphony: if grain duration is 80 ms and density is 100 grains per second, up to 8 grains overlap at any instant. Doubling either number doubles the CPU cost — the load grows as $O(d \cdot \tau)$ where $d$ is density (grains/s) and $\tau$ is grain duration in seconds.
• Real-time constraints: audio hardware demands samples every ~20 ms (at 44.1 kHz with a 512-sample buffer). Miss a deadline and you hear a click or dropout. This is a hard real-time scheduling problem — the same family studied in operating systems and by complexity theorists asking whether feasible schedules always exist for a given task set.
• Parameter space: pitch scatter, position jitter, envelope shape, and grain waveform together define a combinatorial texture space. Searching it for a desired sound is not unlike constraint satisfaction — you have many interacting parameters and a perceptual target that is hard to encode formally.

Curtis Roads, who wrote the definitive textbook on granular synthesis in 2001, estimated that a single second of dense granular texture can require scheduling and mixing tens of thousands of individual grain events. Modern implementations use priority queues, lock-free ring buffers, and SIMD vector math to keep up — the same data-structure toolkit that powers scheduling algorithms in operating systems.

Granular thinking shows up far beyond the synthesizer rack:

• Time-stretching without pitch shift: phase vocoders and granular engines are the technology behind slowing a vocal performance to half speed for a film score without making the singer sound like a chipmunk in reverse. Every digital audio workstation ships one.
• Speech and audio codecs: modern compressed audio (AAC, Opus) models sound as overlapping short-time frames — conceptually granular. Each frame is encoded independently, giving the decoder just enough information to reconstruct the cloud.
• Environmental sound design: video game engines use granular techniques to generate endless variation from a single footstep sample — scatter its pitch, timing, and amplitude slightly and the same one-second clip sounds different every step.
• Scientific signal processing: seismologists, bioacousticians, and radar engineers analyze signals by decomposing them into short-time windows (spectrograms) — Gabor's original insight in disguise.
• Machine listening: mel-frequency cepstral coefficients (MFCCs), the backbone of speech recognition, are computed over short overlapping frames — the same grain-like slices granular synthesis uses to build sound.

Dennis Gabor's acoustic quanta were a physicist's abstraction. Curtis Roads turned them into a composer's tool. Today every digital audio workstation, speech codec, and game engine quietly schedules grain-like frames — because the ear, it turns out, hears statistics, not events.

The deeper lesson is one of representation: by choosing the right unit (the grain, the frame, the quantum), you can decompose almost any signal into something manageable, schedulable, and recombineable. That idea connects acoustic physics to the algorithmic heart of scheduling and even touches the information-theoretic limits explored in data compression. Sound, like computation, is richer when you look closely at its smallest pieces.