What a grain is
Granular synthesis takes a sound apart and reassembles it from pieces. Each piece is a short excerpt called a grain, taken from the source, multiplied by a window, overlapped with its neighbours and summed. The window is what separates the technique from simple looping. A rectangular window starts and ends at full amplitude and produces the clicking that made granular synthesis famous in the wrong way. A raised cosine or Hann window multiplies the grain to zero at both edges, so nothing is discontinuous.
The second thing a grain must avoid is repeating in phase. Two identical grains arriving together add perfectly and produce a comb filter that sweeps as the overlap moves. This is why grain position is randomised even in plugins that expose it as a dial: a few milliseconds of jitter destroys the periodicity that makes the artefact audible.
| Parameter | Working range | What it does |
|---|---|---|
| Grain size | 20 to 300 ms | Sets resolution. Small is noisy, large is smeared |
| Grain rate | 0.5x to 4x | How many grains per second. Sets the texture |
| Position jitter | 0 to 20 ms | Breaks up comb filtering |
| Position offset | Manual or random | Chooses which part of the source dominates |
| Reverse probability | 0 to 100 percent | Adds the synthetic, reversed quality |
Those controls are the whole instrument. Envelope tables, freeze modes and spectrum displays in commercial plugins are ways of drawing motion across those same five values.
Size, rate and position in practice
Grain size sets the resolution more than anything else. At 20 ms a grain is near the limit of pitch perception and the result reads as noise rather than as a note. At 300 ms individual grains become audible events and the texture turns into a rattle or a swarm. Sweep the size while a sustained chord holds and the same source moves from vapour to gravel without its pitch moving at all, which is the most useful single gesture in the technique.
- 20 to 40 ms for noise beds, breath, wind and air.
- 60 to 120 ms for pitched pads and flute-like fragments.
- 150 ms and above for smears, transitions and background wash.
- Grain rate above about 2x on a short grain gives a granular crackle that works well under percussion.
Position control is what makes granular an instrument rather than an effect. Locking position to a single point freezes one region of the source; sweeping it across the source walks the texture through the recording; randomising it densely destroys any sense of location and produces pure material with no memory of where it came from.
Stretching time without moving pitch
The famous application is time-stretching, where grain rate and playback rate run in opposition so that perceived time stretches while the perceived frequency does not. It works well on sustained material and badly on transients. A kick in the source becomes a soft click spread across the stretched window, because the attack is shorter than the grain and a grain cannot preserve something shorter than itself.
That failure is why phase vocoders took the popular market. A vocoder analyses into short overlapping frames, so an attack survives as long as the analysis window can hold it, and it degrades far more gracefully on percussive material. Its own artefact is the opposite one: phasiness, a hollow chorused quality on sustained sounds. The two methods fail in complementary places.
The practical threshold is simple. If the material has an attack shorter than roughly 50 ms, use a vocoder. If the material is essentially continuous, use granular. Around the crossover, listen both ways before you commit, and keep the dry signal reachable for comparison, because granular processing makes it very easy to build something impressive and unusable.
Freezing and the limits you cannot pass
Freeze mode holds a single window, usually one grain, indefinitely. Holding it by looping is exactly the artefact described above and sounds like a jammed tape loop. Spectral freeze avoids that by keeping phase relationships while resampling the spectrum, which is why it reads as bowed glass rather than as a stuck record.
Some things still cannot be recovered. Freeze a transient and you get the aftermath, not the hit. Reverse every grain and the result is obviously synthetic, which is sometimes precisely what you want. Beyond that there is a trade with no free side: clarity of the original against control over the result, and the more control you take the further the material moves from the source.
The bottom line
Granular is the only method that changes the texture of a sound without changing its pitch or its length, and that single capability is worth learning properly. Its failures are predictable enough to route around: use it on continuous material and let a phase vocoder handle anything with an attack.
Stretch a phrase and keep the pitch
432Hz MASTER moves the whole production by one ratio, so a stretched phrase stays at the same pitch relationship to everything else in the session.
Frequently asked questions
How small should a grain be?
Around 20 to 50 ms for noise, air and texture. Around 60 to 120 ms for pitched material. Above 150 ms individual grains become audible as separate events and the texture becomes a rattle.
Why does granular stretching sound better than a phase vocoder on some sounds?
Because granular reconstruction resynthesises from scratch rather than repairing the original waveform. On sustained pads and drones that produces a cleaner, fuller result. On drums and vocals a phase vocoder wins, because its short analysis frames preserve attacks that a long grain necessarily blurs.
Can granular synthesis be free of artefacts?
No. The main artefacts are phasey, metallic ringing from looped windows and loss of transients, and they all come from the same cause: a grain is a fixed window of audio and the source inside it moves. Position jitter of a few milliseconds removes most of the ringing.
What sample rate should I stretch at?
Work at your session rate. Stretching after resampling to a lower rate throws away the transients before the algorithm has a chance to analyse them, and the artefacts become permanent rather than recoverable.