Adaptive Grain Cloud Synthesis — User Guide

What this does

Adaptive Grain Cloud Synthesis resynthesizes the selected Sound as a cloud of short, overlapping grains. The source is converted to mono, grains are read from positions that follow the source from beginning to end (or remain fixed for the Spectral Freeze preset), and each grain is placed on an exact sample-quantized output event grid.

The cloud can vary in density, overlap, source-position scatter, pitch scatter, grain duration, and direction. When Adaptive_duration is enabled, grain duration responds to the source's spectral-centroid profile: relatively darker regions produce longer grains and relatively brighter regions produce shorter grains. This adaptation affects duration only; pitch scatter remains controlled separately by Pitch_scatter_semitones.

What is a grain cloud?

A grain is a short excerpt of sound. A grain cloud is a collection of many such excerpts whose start times, source positions, durations, pitch shifts, and directions can vary. When grains overlap densely, the individual fragments merge perceptually into a continuous texture; with lower density or larger spacing, the individual events remain more audible.

Quick start

1. In Praat, select exactly one Sound object.
2. Run Adaptive_Grain_Cloud_Synthesis.praat.
3. Choose a preset, or keep Custom.
4. Set Grain_size_ms, Grain_overlap, and Density.
5. Use Pitch_scatter_semitones for random per-grain pitch variation and Position_scatter for random variation of the source read position.
6. Enable Adaptive_duration if you want spectral centroid to control grain duration.
7. Enable Reverse_random if approximately half of the grains should be reversed.
8. Set Output_duration_factor to define the requested cloud duration relative to the source.
9. Click OK. The output is created as <source>_grainCloud.

Presets

Choosing a named preset overrides the grain, scatter, processing, and output-duration values shown below. The preset does not change Draw_visualization or Play_result.

Adaptive duration

When Adaptive_duration is enabled, the script divides the source into a small number of analysis windows: 20 for files of at least 1 second, 10 for files shorter than 1 second, and 5 for files shorter than 0.25 seconds. Each window is Hanning-windowed, transformed to a Spectrum, and measured with Get centre of gravity: 2.

The mapping is relative to the spectral-centroid range of the selected source, not to a fixed brightness threshold such as 2000 Hz. A source with little centroid variation uses the neutral 1.0× duration multiplier.

Scheduling & rendering

Event rate

Density therefore changes the actual event rate. At density 2, the requested event hop is half the base hop; at density 0.5 it is twice the base hop.

Sample-quantized scheduler

The requested event hop is rounded to an integer number of samples. A minimum event hop of 1 ms is enforced. The scheduler may further increase the hop when necessary to respect two internal safety limits: at most 10,000 grains and at most 24 tracks. If this happens, the Info window reports that the effective density was reduced.

Source traversal and position scatter

Outside Spectral Freeze, the nominal source read position travels linearly from the beginning toward the last valid base-grain start as output time advances. Position_scatter adds Gaussian displacement to that source read position:

The result is clamped to the valid source range. Position scatter does not jitter output event times.

Pitch scatter

Each grain receives a Gaussian random pitch shift with standard deviation Pitch_scatter_semitones. The draw is bounded to ±3σ and also to an absolute maximum of ±24 semitones. Shifts whose absolute value is 0.1 semitone or less are left unchanged.

Pitch shifting is implemented by overriding the grain sampling frequency by 2^(shift/12) and resampling back to the source rate. Because this method changes the grain's sample count, pitch shifting also changes grain duration.

Tracks and mixing

Scheduled grains are interleaved over enough tracks that grains assigned to the same track do not overlap. Each grain is followed by silence to fill one exact track slot, every track is forced to the exact requested output sample count, and all tracks are then summed. Before summing, each track is scaled by 1 / sqrt(number_of_tracks).

Parameters

Visualization

When Draw_visualization is enabled, the Picture window shows four components:

• Source: the mono analysis/render source waveform, shown on a fixed ±1 amplitude scale.
• Grain time map: every scheduled grain is drawn as a mapping from its source interval on the horizontal axis to its output interval on the vertical axis. Blue upward mappings are forward grains; purple downward mappings are reversed grains. A pale reference line shows the nominal linear source traversal, or a vertical line in Spectral Freeze.
• Output: the final mono grain cloud on a fixed ±1 amplitude scale.
• Summary: grain count, base grain size, overlap, track count, requested/effective density, position scatter, pitch σ, duration, adaptive-duration state, and random-reverse state.

The Source and Output waveform panels use the same amplitude scale, so their displayed heights can be compared directly.

Output, limits & behavior

• The source must be at least as long as the base grain size.
• Grain overlap must be between 0 and 0.9.
• Position scatter must be between 0 and 1.
• Pitch scatter must be zero or positive.
• Output duration factor must be positive.
• The scheduler uses a minimum event hop of 1 ms, at most 10,000 grains, and at most 24 tracks.
• The output duration is sample-quantized from round(sourceSamples × Output_duration_factor).
• The final cloud is unconditionally peak-normalized to 0.9 when non-silent.
• A 10 ms fade-in and fade-out are applied; for outputs shorter than 40 ms, each fade is reduced to one quarter of the output duration.
• The output object is named <source>_grainCloud.