Advanced Poisson Synthesis — User Guide

Generates layered granular textures from homogeneous Poisson point processes. Each event triggers a raised-cosine-windowed sinusoidal grain; five synthesis modes shape event rate and frequency fields, while optional stereo processing and statistical QC expose the realized stochastic structure.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel Version: 0.4 reviewed (2026) License: MIT License Repo: Praat AudioTools
Contents:

What this does

Advanced Poisson Synthesis creates a new sound from scratch. For each layer, Praat generates a homogeneous Poisson PointProcess over the requested duration. Every realized point becomes the onset of a sinusoidal grain shaped by a raised-cosine (Hann-type) envelope. The layers are summed, globally faded, optionally converted to stereo, and optionally normalized.

Core timing rule: a layer with rate λ has an expected point count of λT over duration T. Consecutive inter-onset intervals Δt are exponentially distributed; the normalized interval z = λΔt therefore follows Exp(1), whose mean and coefficient of variation are both 1. The script measures these quantities from the actual generated realization.

Important behavior:

Quick start

  1. Run Advanced_Poisson_Synthesis.praat in Praat.
  2. Choose a Preset, or leave Custom to use the form values directly.
  3. For a first test, try Standard Three Layer.
  4. Leave Normalize output, Draw visualization, and Play result enabled if you want the normalized sound, QC figure, and immediate playback.
  5. Use a positive Random seed when you need the same stochastic realization again.
Randomize parameters: when enabled, the script can jitter layer rates and randomize grain frequency, duration, and phase. Turning it off fixes those parameters to deterministic values within each mode; the Poisson event times themselves are still random unless a positive seed is supplied.

Poisson model and grain construction

Layer timing

For each layer: expected point count = λ × duration Praat: Create Poisson process(..., λ) For consecutive points: Δt = t[i] - t[i-1] z = λ × Δt Poisson prediction: z ~ Exp(1)

The script pools normalized inter-onset intervals from all layers for QC. It also compares the total realized point count with the summed expectation using a count z-score:

count z = (realized points - expected points) / sqrt(expected points)

Grain signal

Each rendered point triggers one sinusoidal grain. The carrier starts at the event time and is multiplied by a raised-cosine envelope:

grain(t) = A × sin(2πf(t-t0) + phase) × [1 - cos(2π(t-t0)/D)] / 2 for t0 ≤ t < t0 + D

Requested grain durations are converted from milliseconds to seconds. The minimum is raised to at least eight samples. A grain that reaches beyond the end of the requested output is truncated at the boundary, and a final fragment shorter than four samples is not rendered.

Layer energy balance

Each mode supplies a base amplitude factor, but layer gain is also divided by the square root of the expected overlap λ × mean grain duration. This reduces the extent to which increasing event density also acts as an unintended level control.

Synthesis modes

The modes change the layer-rate rule, frequency field, and a mode amplitude factor. Grain duration and phase follow the global randomization switch.

ModeLayer rateRaw grain-frequency fieldMode factor
Rate Gradient Linear low→high gradient across layers; optional ×0.8–1.2 rate jitter. base ... base + range 1.5
Dense Granular 1.5 × high rate; optional ×0.7–1.3 jitter. base × (0.5 + 0.3×layer) ... + range 1.2
Sparse Atmospheric 0.5 × low rate; optional ×0.6–1.4 jitter. base × (0.3 + 0.4×layer) ... + 0.5×range 2.0
Steady-rate Poisson Pulses Midpoint of low/high rates; optional ×0.9–1.1 jitter. base × layer ... + 0.3×range 1.8
Broad Poisson Scatter With randomization: random low→high rate, then ×0.5–1.5. Without randomization: midpoint rate. Broad field from approximately 0.5×base to 2.5×base + range. 1.5
When Randomize parameters is off: grain frequency and duration use the midpoint of their mode-specific ranges and grain phase is 0. In Broad Poisson Scatter, the layer rate also becomes the low/high midpoint. The onset process remains Poisson-timed.

Frequency fitting and anti-alias guard

The safe upper frequency is 0.45 × Sample_rate_Hz. If a mode's requested raw maximum exceeds that boundary, the entire requested field for that layer is proportionally compressed into the safe band. The QC report counts how many rendered grain frequencies required this fitting step.

Presets

Preset selection occurs before validation. The table below lists the values each preset explicitly overrides.

PresetValues explicitly set
CustomNo preset overrides; form values are used.
Standard Three Layer12 s; base 100 Hz; range 300 Hz; rates 3–15 Hz; 3 layers; grains 100–300 ms; Rate Gradient; Mono.
Dense Cloud10 s; base 150 Hz; range 400 Hz; rates 10–25 Hz; 4 layers; grains 30–80 ms; Dense Granular; Stereo Wide.
Sparse Atmosphere20 s; base 80 Hz; range 500 Hz; rates 1–5 Hz; 3 layers; grains 300–800 ms; Sparse Atmospheric; Rotating; fade 3 s.
Poisson Pulse Field15 s; base 120 Hz; range 200 Hz; rates 5–12 Hz; 4 layers; grains 80–120 ms; Steady-rate Poisson Pulses; Mono; Randomize parameters off.
Broad Scatter12 s; base 100 Hz; range 600 Hz; rates 2–20 Hz; 5 layers; grains 50–300 ms; Broad Poisson Scatter; Stereo Wide.
Shimmering High10 s; base 800 Hz; range 1500 Hz; rates 8–20 Hz; 4 layers; grains 20–60 ms; Dense Granular; Stereo Wide.
Deep Rumble15 s; base 40 Hz; range 80 Hz; rates 2–8 Hz; 3 layers; grains 200–500 ms; Sparse Atmospheric; Rotating; fade 3 s.
Fields retained from the form: presets do not overwrite Sample_rate_Hz, Random_seed, Normalize_output, Draw_visualization, or Play_result. They also retain the user's Randomize_parameters setting except Poisson Pulse Field, which explicitly turns it off. Fade time is retained except in Sparse Atmosphere and Deep Rumble, which set it to 3 seconds.

Parameters

ParameterDefaultBehavior
PresetCustomSelects one of eight configurations; see the preset table above.
Duration_s12Output duration in seconds.
Sample_rate_Hz44100Sampling frequency. Script accepts 4,000–384,000 Hz.
Base_frequency_Hz100Reference frequency used by the selected mode.
Frequency_range_Hz300Mode-dependent width added to the raw frequency field.
Number_of_layers3Independent Poisson layers; clamped to 1–8.
Low_rate_Hz3Lower Poisson-rate control. If entered above High rate, the two values are swapped.
High_rate_Hz15Upper Poisson-rate control; exact use depends on synthesis mode.
Min_grain_duration_ms30Minimum requested grain duration; swapped with Max if necessary and raised to at least eight samples.
Max_grain_duration_ms200Maximum requested grain duration; never lower than the validated minimum.
Synthesis_modeRate GradientSelects the layer-rate and raw frequency rules.
Fade_time_s2Linear global fade-in and fade-out. Limited to at most half the total duration.
Spatial_modeMonoMono, Stereo Wide, or Rotating.
Randomize_parametersYesControls layer-rate jitter and grain frequency/duration/phase randomization. Does not disable Poisson timing.
Random_seed0Positive value = reproducible stochastic stream. 0 = unpredictable stream.
Normalize_outputYesFor a nonzero result, Scale peak: 0.9 performs target peak normalization to 0.9.
Draw_visualizationYesDraws the three-panel Poisson QC figure in Praat's Picture window.
Play_resultYesPlays the generated Sound at the end.

Workload guards

Before synthesis, the script estimates a conservative event workload from duration, layer count, mode, and rate settings. It stops if that estimate exceeds 25,000 grains. During generation it also stops if the realized Poisson point count exceeds 30,000. Grains are accumulated in formula chunks of 25 events rather than by constructing one enormous per-layer formula.

Output and spatial modes

Mono

The summed generator remains one channel. No stereo conversion is performed.

Stereo Wide

The mono generator is copied into two channels and filtered differently before Combine to stereo. At 44.1 kHz the left copy passes approximately 0–4 kHz and the right copy approximately 200 Hz–8 kHz, with a 100 Hz Hann transition. These limits adapt downward when the sample rate requires it.

The two channels therefore differ spectrally; this mode is not a simple pan or duplicated dual-mono output.

Rotating

The mono generator is converted to stereo with complementary equal-power gains:

θ(t) = π/4 × [1 + sin(2π × 0.25 × t)] Left = input × cos(θ) Right = input × sin(θ) cos²(θ) + sin²(θ) = 1

The panning function repeats every 4 seconds and moves smoothly between the channels while maintaining equal-power complementarity.

Fade and normalization order

The global linear fade is applied to the summed mono texture first. Spatial processing comes next. Peak normalization, when enabled, is performed after the final mono/stereo output has been created. Therefore the 0.9 target applies to the final output object.

Output name: the final Sound is named poisson_<preset name>, for example poisson_Standard, poisson_DenseCloud, or poisson_Custom.

Visualization and statistical QC

When Draw_visualization is enabled, the script draws a fixed three-panel research view based on the actual realization generated in that run.

Panel A — Actual Poisson event realization

A time-by-layer raster shows rendered grain onsets as vertical ticks. Each row is one independent layer. For legibility, the display uses a stride when more than about 1,200 rendered grains are available; this does not alter the audio or the stored event statistics.

Panel B — Poisson timing QC

The script measures normalized intervals z = λΔt and plots their density over z = 0…5. The theoretical Exp(1) density exp(-z) is overlaid, together with a vertical marker for the measured normalized-IOI mean. The QC dashboard also reports the measured coefficient of variation; the Exp(1) targets are mean = 1 and CV = 1.

Panel C — Model → measurement

A measured spectrogram of the generated output is shown with the realized grain frequencies and durations overlaid as horizontal trajectories. For stereo output, the script extracts both channels, measures RMS, and displays the stronger channel; ties use the left channel. The overlay is subsampled when more than about 450 grains are available. The spectrogram upper frequency follows the realized grain range, bounded by the script's safe upper frequency.

Bottom QC dashboard

The summary reports:

Compositional context

This tool combines two established computer-music ideas: microsound/granular construction from short sound particles and stochastic control of event placement. Its particular implementation uses Praat's homogeneous Poisson point process to generate onsets, then maps those events to windowed sinusoidal grains whose frequency, duration, phase, density, and stereo treatment are controlled by the selected mode.

In a broader historical context, stochastic procedures became an explicit compositional resource in the work and writings of Iannis Xenakis, while granular and microsound techniques were developed into a substantial compositional practice documented by Curtis Roads. Here, the historical connection is the use of probability and sound particles as compositional control structures; the audible result is determined by the specific Poisson-and-grain model implemented in this script.

Further Reading