Subtractive synthesis starts from a waveform and carves it down. Granular synthesis starts from a recording — any recording — and blows it apart into hundreds of tiny fragments, then lets you decide how those fragments get put back together. It’s one of the more disorienting techniques to learn precisely because there’s no single “the sound” to point to; the sound is whatever you do with the grains.
What a grain actually is
A grain is a very short slice of audio, typically somewhere between 1 and 100 milliseconds — too brief to be heard as a distinct note on its own, but long enough to carry a fragment of the original sound’s timbre. A granular synthesizer takes a source (a sample, a live input, a field recording) and continuously carves grains out of it, then replays them according to a handful of core parameters: grain size (how long each fragment is), density (how many grains play per second, and how much they overlap), position (where in the source material each grain is pulled from), and pitch (which can be shifted independently of playback speed, since each grain is just being replayed, not stretched). Randomize position and density enough, and a granular engine turns a single clean recording into a shifting cloud of texture; keep them tight and ordered, and it can sound almost like the source material played normally, just subtly smeared.
A reasonable first project
Load a short vocal or instrumental sample into any granular tool (Ableton’s Grain Delay, a Max/MSP patch, or a dedicated granular synth like Output’s Portal all work) and start with grain size around 30-50ms and low density, so individual grains are still audible as discrete events. From there, try three moves in isolation: sweep the position parameter slowly across the sample to hear how it “reads through” different parts of the source, randomize pitch per grain to turn a clean tone into a shimmering cluster, and crank density until individual grains blur into a continuous texture. Doing each in isolation, rather than tweaking everything at once, is the fastest way to build an intuition for what each parameter is actually doing to the sound.