Formant Grain Texture — User Guide
Stochastic source–filter granular synthesis: short harmonic/noise grains arrive as a Poisson event field, are assigned to one of five synthetic vowel banks, filtered through three-resonance FormantGrids, and mixed as a mono or stereo vowel cloud.
What this does
Formant Grain Texture combines granular event scheduling with a small source–filter synthesis model. Every event creates a short excitation grain. The grain is assigned to one of five vowel labels — a, i, u, e, o — and is added to that vowel's excitation bus. Each bus is then filtered through its own static three-resonance FormantGrid.
Poisson onset times
↓
Hann-windowed harmonic/noise grains
↓
5 vowel excitation buses
↓
5 static F1/F2/F3 FormantGrids
↓
mono or stereo vowel-bus mix
↓
edge fade → optional peak normalization
This architecture is important: the formants are filter resonances, not three sine oscillators added to each grain. Because all grains assigned to one vowel use the same static linear filter, filtering their summed bus is mathematically equivalent to filtering those grains individually and summing the results afterward.
The vowel frequencies are synthesis targets. They are not measured from an input Sound, and the script requires no selected Sound object.
Poisson grain field
Grain density is the expected event rate in grains per second. The script generates chronological inter-onset intervals from an exponential distribution:
inter-onset interval ~ Exponential(rate = Grain density)
Therefore the expected grain count is:
expected grains = Grain density × Duration
The actual count is stochastic. Two runs with Random seed = 0 normally produce different onset schedules and different actual counts.
The script refuses settings whose expected count exceeds 10,000 grains and has a second guard at 12,000 actual grains for an unusually large stochastic realization.
Grain duration is drawn uniformly between the Min and Max grain-duration controls, then shortened if necessary so that a grain beginning near the end never extends beyond the requested output duration.
Per-grain level includes expected-overlap compensation. The nominal gain decreases approximately with the square root of expected overlap, so increasing density primarily changes occupancy and textural mass instead of functioning as an unintended loudness envelope.
Excitation grains
Every grain receives a full Hann envelope and a random starting phase.
Voiced mode: Base frequency > 0
The excitation is a harmonic series whose amplitudes fall as 1/h. The harmonic stack is RMS-normalized before breath noise is mixed in. Harmonics above the practical upper limit of 0.45 × sample rate are omitted.
F0 jitter perturbs each grain independently around Base frequency by a uniform percentage range. For example, 10% jitter draws F0 between approximately 0.9× and 1.1× the base value, subject to the 20 Hz lower limit and practical sampling headroom.
Breath mix
Breath mix is treated as an approximate power fraction:
source = sqrt(1 - breath) × harmonic
+ sqrt(breath) × Gaussian noise
Thus 0 means harmonic excitation only, while 1 means noise only.
Whisper mode: Base frequency = 0
When Base frequency is exactly zero, no harmonic series is generated. Each grain uses Gaussian-noise excitation only. The Whispered Back-Vowel Cloud preset uses this mode.
Random seed
A positive Random seed reproduces the complete stochastic synthesis path: distorted-bank perturbations, Poisson schedule, vowel assignments, durations, phases, F0 jitter, and rendered noise. Seed 0 uses the current unpredictable random state. After a fixed-seed render, Praat's unpredictable random initialization is restored.
Five synthetic vowel banks
| Bank | F1 | F2 | F3 |
|---|---|---|---|
| a | 730 Hz | 1090 Hz | 2440 Hz |
| i | 270 Hz | 2290 Hz | 3010 Hz |
| u | 300 Hz | 870 Hz | 2240 Hz |
| e | 530 Hz | 1840 Hz | 2480 Hz |
| o | 570 Hz | 840 Hz | 2410 Hz |
The three synthesis bandwidths are 90, 130, and 180 Hz, multiplied by Formant bandwidth scale. They remain constant over time within each vowel bus.
At unusually low sample rates or in Distorted Vowel Bank, the script protects sampling headroom and enforces the ordering F1 < F2 < F3. Distorted Vowel Bank randomly perturbs all five banks once per run before those safeguards are applied.
After filtering with Filter (no scale), each vowel bus receives one global scalar that matches its filtered RMS back to the RMS of its excitation bus. This compensates the potentially large absolute gain of the all-pole filter while preserving the bus's filtered spectral shape and its internal grain-to-grain dynamics.
Presets
Presets mainly set grain density, F0, duration range, filter bandwidth, source harmonics, breath mix, and jitter. Unless stated below, they leave the user-selected Duration, Sample rate, and Spatial mode unchanged. Normalize, Draw, Play, Random seed, and Edge fade are not preset overrides.
| Preset | Density | F0 | Grain ms | BW scale | Harmonics | Breath | Jitter | Vowel selection |
|---|---|---|---|---|---|---|---|---|
| Balanced Vowel Cloud | 20/s | 120 Hz | 50–120 | 1.00 | 12 | 0.04 | 8% | Uniform a/i/u/e/o |
| Whispered Back-Vowel Cloud | 15/s | 0 | 80–180 | 1.35 | 12* | 1.00 | 0% | a 40%, u 30%, o 30% |
| Alternating Front Vowels | 35/s | 80 Hz | 30–60 | 0.65 | 16 | 0.01 | 3% | Random 50/50 i/e; not deterministic alternation |
| Distorted Vowel Bank | 30/s | 140 Hz | 40–100 | 0.80 | 15 | 0.06 | 14% | Uniform; bank targets are randomly perturbed once per run |
| Sparse Low Open Vowels | 12/s | 90 Hz | 120–300 | 0.90 | 12 | 0.02 | 5% | a 40%, u 30%, o 30%; Duration becomes 8 s |
| High Dense Vowel Scatter | 45/s | 280 Hz | 25–60 | 0.80 | 18 | 0.04 | 12% | Uniform a/i/u/e/o |
| Sparse Breathy Back Vowels | 8/s | 160 Hz | 150–350 | 1.25 | 10 | 0.18 | 10% | o 60%, i 40%; forces Rotating Vowel Buses |
| Dense Narrow Vowel Storm | 60/s | 110 Hz | 15–40 | 0.65 | 20 | 0.08 | 18% | Uniform; forces Wide Vowel Field |
* In whisper mode the harmonic-count setting is retained but unused because F0 = 0.
Spatial modes
Spatialization is applied to the five filtered vowel buses, not independently to every grain. Each vowel therefore has a coherent spatial identity.
| Mode | Implementation |
|---|---|
| Mono | The five filtered vowel buses are summed directly to one channel. |
| Stereo Vowel Spread | Fixed equal-power bus positions: a=0.08, i=0.29, u=0.50, e=0.71, o=0.92. |
| Rotating Vowel Buses | Each vowel bus follows a 0.10 Hz sinusoidal pan trajectory with a different 72° phase offset. Pan moves between approximately 0.04 and 0.96. |
| Wide Vowel Field | Fixed wide positions: a=0.03, i=0.97, u=0.14, e=0.86, o=0.50. |
For stereo modes the gains are equal-power-style sqrt(1-p) and sqrt(p). The current version does not use complementary left/right spectral coloration.
Rendering and output level
Excitation buses are built in chunks of at most 0.5 seconds. The script uses Formula (part) only over the current chunk. A grain crossing a chunk boundary keeps the same onset-relative age, harmonic phase, duration, and Hann envelope on both sides of that boundary.
Chunks are only an implementation strategy: they do not introduce output overlap, crossfades, or timeline stretching.
Edge fade and normalization
A final linear edge fade is applied to the complete mono/stereo result. Its actual length is min(Edge fade, 20% of Duration).
If Normalize output is enabled, every non-zero result receives Scale peak: 0.90. This is target peak normalization, not an attenuate-only ceiling. With normalization disabled, the raw post-filter/post-spatial level is retained.
| Output property | Behavior |
|---|---|
| Duration | Requested duration, except Sparse Low Open Vowels sets 8.0 s. |
| Sample rate | Direct synthesis at the selected rate, 8–192 kHz. |
| Channels | Mono in Mono mode; stereo in the other three modes. |
| Output name | FormantGrain_<preset name> with spaces replaced by underscores. |
Visualization and QC
| Panel | What it actually shows |
|---|---|
| A — Stochastic Event Field | Actual onset-to-end grain segments arranged on five rows by assigned vowel. |
| B — Synthesis Formants | Actual F1/F2/F3 targets assigned to sampled rendered grains. |
| C — Model → Measurement | Measured final-output spectrogram with sampled actual F1/F2/F3 grain guides overlaid. |
| D — Measured Output | Measured final waveform. For stereo output the channel with the higher whole-file RMS is used as the representative display channel. |
The QC area reports expected and actual event count, realized density, seed state, vowel occupancy, mean grain duration, source/filter settings, spatial mode, and pre/final output levels.
Further Reading
- Praat Manual — Source-filter synthesis / FormantGrid. Direct documentation for the FormantGrid source–filter mechanism used by this script.
- Roads, C. (2002). Microsound. MIT Press. Direct background on granular and microsound techniques built from short sound particles.