Rich Formant Grains — User Guide

What this does

Rich Formant Grains is a source–filter granular generator. It does not synthesize F1, F2, and F3 as oscillators. Instead, each grain first creates a broadband harmonic/noise excitation, and that excitation is then passed through a three-formant resonant filter bank.

The complete process is:

fixed-count grain plan
    → harmonic/noise excitation × half-sine grain window
    → one of 15 F1–F3 resonance buses
    → per-bus RMS compensation
    → sum of all buses
    → edge fades
    → spatial mode
    → selected level mode

Quick start

1. Run Rich_Formant_Grains.praat. No input Sound is required.
2. Choose Custom or one of the eight presets.
3. Set Duration, Grain density, Base frequency, Breath noise mix, and Spatial mode.
4. Enable Edit details to change sample rate, harmonic count, formant bandwidths, grain-duration range, level mode, output peak/ceiling, and random seed.
5. Run the script. The complete stochastic grain plan is generated first, then rendered through the fifteen resonance buses.

Grain plan

Grain density does not drive a Poisson process. The script first fixes the number of grains:

totalGrains = round(Duration × GrainDensity)
realizedDensity = totalGrains / Duration

Because the count is rounded to an integer, the realized density can differ slightly from the requested density.

Duration and onset

Each grain duration is drawn uniformly between Minimum grain duration and Maximum grain duration, after both limits are capped by the total output duration.

Duration is drawn before onset. Onset is then chosen uniformly from:

0 … Duration - grainDuration

Consequently every grain fits completely inside the output. There is no end-of-file truncation of randomly placed grains.

Grain onsets are independent and can overlap freely. With high density or long grains, substantial temporal overlap is expected.

Vowel and resonance-bus assignment

Five vowel templates are available: a, i, u, e, o. Each vowel has three resonance variants, giving fifteen buses in total.

For most presets the vowel is chosen uniformly from the five templates. Whisper Choir and Gregorian Chant use a weighted population of a/u/o = 40%/30%/30%. Ghost Voices chooses only i or u with equal probability.

After the vowel is chosen, one of its three resonance variants is selected uniformly for that grain.

Excitation source

Each grain receives an F0 drawn independently from:

grainF0 = BaseFrequency × U(0.90, 1.10)

Thus Base frequency is the center of a ±10% F0 spread. Base frequency may also be set to 0, in which case the excitation is noise-only.

Harmonic component

For nonzero F0, the harmonic source sums integer harmonics up to the user-selected Source harmonics limit, additionally truncated so that harmonic frequencies remain below Nyquist − 100 Hz:

harmonic amplitude ∝ 1 / h^0.85

Every grain resets the harmonic phases at its own onset because each sinusoid uses local grain time.

Breath-noise mixture

The excitation before windowing is:

(1 - n) × harmonicSource + n × GaussianNoise

where n is Breath noise mix in the range 0–1. At 1.0 the source is pure Gaussian noise.

Grain window and amplitude

Each grain is multiplied by a half-sine window:

w(τ) = sin(π τ / D),   0 ≤ τ ≤ D

The window starts and ends at zero. The grain amplitude also includes a small random factor of 0.85–1.15 and an inverse-square-root density compensation:

grainAmp = ampShape × 0.40 / sqrt(max(12, density)) × U(0.85,1.15)

The preset-specific ampShape is 1.0 except Whisper Choir = 0.8, Robotic Speech = 1.2, and Ghost Voices = 0.7.

F1–F3 resonance banks

The five base resonance templates are:

Three variants per vowel

For ordinary presets, each vowel receives three nearby fixed variants. Variant 2 uses the base values; variants 1 and 3 apply small downward/upward multipliers, with progressively smaller spread from F1 to F3.

Alien Language is different: its three variants are randomly generated at the start of each run from much wider independent multiplicative ranges. Once created, those fifteen bus resonances remain fixed for that realization.

Bandwidths

Nominal bandwidths are:

F1 BW = 90  × bandwidthScale
F2 BW = 130 × bandwidthScale
F3 BW = 180 × bandwidthScale

The resonance frequencies are kept ordered and constrained below Nyquist.

Filtering and RMS compensation

Each of the fifteen excitation buses is filtered by a Praat FormantGrid with three constant resonances. Filtering uses Filter (no scale).

After filtering, each bus is multiplied by one scalar gain so that its output RMS matches its pre-filter source RMS:

busGain = sourceRMS / filteredRMS

This compensation prevents the raw gain of the all-pole formant filter from becoming an unintended dynamic control. It does not flatten the formant spectrum; the spectral shaping remains.

Spatial modes

The three stereo modes are created from the already summed mono source–filter result; individual grains are not independently panned.

Level modes

After spatial processing the script measures the pre-level peak and RMS, then applies one of three modes:

Controls

Main page

Details page

Presets

Presets override Grain density, Base frequency, Source harmonics, Breath noise mix, Formant bandwidth scale, grain-duration range, and in some cases the internal amplitude-shape factor. They do not override Duration, Spatial mode, Sample rate, Output mode, Output peak/ceiling, or Random seed.

The preset names describe intended timbral directions only; they are not analyses or models of actual speech, chant, infant vocalization, or language.

Sampling and safety

Sample rate must be at least 8000 Hz, and the minimum grain duration must span at least eight samples.

If Base frequency is above zero, the upper edge of its +10% F0 spread must remain below Nyquist − 100 Hz. Individual harmonic counts are additionally truncated to the same safety margin.

All F1–F3 resonance variants are kept ordered and constrained below Nyquist. At low sample rates this can alter the programmed template values to keep the filters valid.

Runtime guard

Version 1.1.1 limits the fixed grain plan to:

maximum grains = 12,000

If round(Duration × GrainDensity) exceeds this limit, the script stops before allocating the grain plan or beginning synthesis.

Randomness and reproducibility

Random choices include grain duration, onset, vowel selection, resonance variant, per-grain F0, amplitude variation, Gaussian excitation noise, and the wide resonance variants used by Alien Language.

Random seed = 0 initializes Praat unpredictably. A positive seed makes the realization repeatable for the same settings.

The script consumes one random integer for an internal temporary-object identifier after seeding; this consumption is deterministic under a fixed seed, so repeatability is preserved.

After a fixed-seed run, Praat's global RNG is restored to an unpredictable state.

Visualization and QC

The QC summary reports grain count, requested/realized density, duration mean/range, mean source F0, noise mix, mean F1/F2/F3, spatial mode, level mode, Nyquist frequency, final peak/RMS, and pre-level peak.

Output