GRM-Style Resonator — User Guide

A stereo resonant filter-bank processor that extracts selected frequency bands, rings each band with a tuned or fixed delay-line resonator, shapes the bank spectrally, and distributes the bands across a stereo field.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel Version: 1.5.1 (2026) License: MIT License Repo: Praat AudioTools

Contents:

What this does Quick start Signal flow Presets Parameters Resonator models Stereo architecture Output behavior Visualization

What this does

GRM-Style Resonator builds a bank of narrow or moderate-width frequency bands from the selected Sound. Each band is multiplied by a frequency-dependent gain, passed through a delay-line resonator, assigned left/right gains, and added to a stereo wet bus. The wet bank is then blended with the source according to dryWet.

The center frequencies can come from a manual list, a harmonic series, or a list of inharmonic ratios. The delay can follow each band center or use one fixed manual delay for every band. Two bounded resonance models are available: a finite echo train and a single recursive feedback comb.

About “GRM-Style”: the name describes an aesthetic connection to resonant spectral transformation associated with GRM Tools. This script is not a reimplementation of the GRM Tools Reson algorithm. The present processor uses Praat Hann band-pass filters followed by delay-line resonators.

Quick start

Select exactly one mono or stereo Sound object in Praat.
Run GRM-Style_Resonator.praat.
Choose a preset, or use Custom to define the frequency bank.
Choose Finite echoes or Single feedback comb.
Set the bank gain, spectral gain profile, stereo architecture, and dryWet.
Choose the output-level mode. Safety ceiling is the default.
Run the script. The result is named <source>_GRMstyle_<preset>.

Starting point: Harmonic Organ is useful for a clearly pitched resonant bank; Inharmonic Metallic and Sparse Bells emphasize non-harmonic ringing; Formant-ish concentrates energy around a fixed set of broad resonance regions.

Signal flow

INPUT SOUND ↓ optional zero tail ↓ frequency-bank generation ↓ Hann band-pass per center frequency ↓ per-band gain profile ↓ finite echo train OR single feedback comb ↓ per-band stereo coefficients ↓ sum to stereo wet bus ↓ dry / wet blend ↓ optional safety ceiling or peak normalization ↓ OUTPUT SOUND

The source is not pitch-tracked. When tuned delay is enabled, the delay is calculated directly from each requested band center frequency.

Presets

Presets override the bank/tuning parameters listed below, but they do not change the selected resonator model, Dry/Wet value, output-level mode, visualization, playback, tail length, random seed, or final ceiling.

Preset	Frequency bank	Bandwidth	Delay / resonance settings	Gain / pan
Custom	Uses the form fields	80 Hz default	Tuned delay; 3 iterations; decay 0.60 by default	+6 dB; Dampen Highs; Wide Spread by default
Harmonic (organ-like)	8 bands: 110 × 1…8 Hz	100 Hz	Tuned delay; 20 iterations; decay 0.85	Dampen Highs; Stereo Spread
Inharmonic metallic	200 Hz × 1, 1.59, 2.14, 2.76, 3.41, 4.07	40 Hz	Tuned delay; 10 iterations; decay 0.70	Flat; Wide Spread
Formant-ish	500, 1500, 2500, 3500, 4500 Hz	200 Hz	Fixed 6 ms delay; 2 iterations; decay 0.40	Dampen Highs; V Shape
Sparse bells	287, 645, 1203, 2156 Hz	35 Hz	Tuned delay; 40 iterations; decay 0.95	+12 dB; Dampen Highs; Random

Resonator model remains user-controlled: the preset values for ringIterations matter only when Finite echoes is selected. In Single feedback comb mode, ringIterations is intentionally ignored.

Parameters

Frequency-bank controls

Control	Default	Meaning
numBands	6	Maximum number of bands used. A manual or ratio list is truncated to this count.
tuningMode	Manual list	Manual center frequencies, harmonic multiples of `baseFreqHz`, or `baseFreqHz × ratio`.
manualFrequencies	300 520 890 1440 2330 3770	Space-separated center frequencies in Hz.
baseFreqHz	110	Base frequency for harmonic and inharmonic-ratio modes.
ratioList	1 1.41 1.89 2.37 2.98 3.56	Space-separated multipliers used by Inharmonic ratios mode.
bandwidthHz	80 Hz	Nominal pass-band width around each center frequency.

Every center frequency must lie above 20 Hz and below Nyquist − 20 Hz. Invalid centers stop the script rather than being silently moved. At the spectrum edges, the actual pass-band limits are constrained to the same usable range.

Band filtering

Each band is extracted with Praat’s Filter (pass Hann band). Transition smoothing follows the actual band width:

smoothing = max(10 Hz, min(100 Hz, actual_bandwidth / 2))

The pale transition regions and darker pass regions are shown separately in the visualization.

Gain profile

Profile	Behavior
Flat (Equal)	Every band receives the linear gain corresponding to `gainDB`.
Dampen Highs	Gain falls from full base gain at the lowest center to 20% of that gain at the highest center.
Boost Highs	Gain rises from 20% of base gain at the lowest center to full base gain at the highest center.

The tilt position is calculated from log frequency, not from the order in which frequencies were entered. An unsorted manual list therefore still receives a frequency-correct tilt.

Advanced settings inside the script

Setting	Default	Meaning
tune_Delay_To_Pitch	1	Use a delay derived from each band center. Despite the historical variable name, this is frequency tuning, not pitch analysis.
manual_Ring_Delay_Ms	8 ms	One fixed delay for all bands when tuned delay is disabled.
tail_ms	0	Append silence before processing so resonance can continue beyond the original ending. Output duration grows by this amount.
random_seed	0	0 gives unseeded Random panning; a non-zero integer makes the random band positions reproducible.
finalPeak	0.99	Target used by Safety ceiling and Peak normalize.
keep_individual_bands	0	When enabled, retains the generated per-band Sound objects after the final mix.

Resonator models

Frequency-tuned integer delay

With tuned delay enabled, each band uses:

D = round(sampleRate / centerFrequency) actual tuned frequency = sampleRate / D

The delay is therefore quantized to whole samples. The actual tuned frequency may differ from the requested band center, especially at high frequencies or low sample rates. The Info window reports the actual frequency and tuning error in cents for every band. This is frequency-derived delay tuning; the script performs no F0 or pitch analysis of the input.

Finite echoes (N taps, bounded)

This model adds delayed copies of the original filtered band:

y[n] = x[n] + a x[n-D] + a² x[n-2D] + … + aᴺ x[n-ND]

ringIterations sets N. ringDecay is a and must be at least 0 and below 1. Because every tap reads the same original band rather than the previous processed result, this is a finite, bounded echo train.

Single feedback comb (decay only)

This model uses one recursive delay:

y[n] = x[n] + a y[n-D]

ringDecay controls the feedback. ringIterations is not used in this mode. A decay of 0 produces no added resonance; values approaching 1 create progressively longer and stronger ringing.

Level: neither resonator model contains an internal peak normalization or hidden wet-bus limiter. High band gain, high decay, multiple overlapping bands, and strong wet mix can therefore produce peaks above 1.0 in Natural level mode.

Stereo architecture

The wet bank is always a two-channel stereo bus. Mono and stereo inputs are handled differently:

Mono input: each resonant band is fanned to left and right according to its panning coefficients.
Stereo input: the left and right source channels are filtered and resonated independently. The panning coefficients act as per-band side gains; the script does not fold the source to mono and does not move left-channel material into the right channel or vice versa.
More than two channels: processing is refused because the wet architecture has no defined placement for additional channels.

Architecture	Left / right coefficients per band
All Center (dual mono)	1.0 / 1.0. On stereo input this preserves the two sides rather than summing them.
Stereo Spread (Alternate)	Alternates 1.0/0.3 and 0.3/1.0.
Wide Spread (Hard L/R)	Alternates 1.0/0.0 and 0.0/1.0.
Left Heavy	1.0 / 0.3 for every band.
Right Heavy	0.3 / 1.0 for every band.
V Shape (Outside In)	Moves coefficients toward the opposite side around the middle of the band list, then outward again.
Random	For each band, R is random in 0…1 and L = 1 − R.

Dry/Wet and output behavior

dryWet ranges from 0 to 1. At values above 0, the output is a stereo blend of dry and wet. A mono dry source is converted to stereo before this blend.

True bypass: when dryWet = 0, the script copies the working source directly, preserves its original channel count, and skips the output-level stage. With tail_ms = 0, this is a sample-preserving bypass. If a tail has been requested, the bypass includes the appended silence and is therefore longer than the original.

Output level modes

Mode	Behavior
Natural level	No level correction. The natural result of band gain, resonance, panning and mixing is retained.
Safety ceiling (default)	If the output peak exceeds `finalPeak` (0.99 by default), the whole result is attenuated to that ceiling. Quieter results are left untouched.
Peak normalize	Any non-silent result is scaled so its peak equals `finalPeak`, whether this requires attenuation or amplification.

If Natural level produces a peak above 1.0, the Sound object is left unchanged and the Info window warns about integer-PCM clipping. Playback uses a temporary peak-scaled copy so audition does not require altering the stored output.

Time, duration, and naming

The source start time (xmin) is restored to the output.
With tail_ms = 0, duration is preserved.
With a positive tail, the output is longer by exactly that appended silence duration.
Sample rate is preserved.
Output name: <source>_GRMstyle_<presetName>.

Visualization

When Draw_visualization is enabled, the script creates a suite-standard 8-inch diagnostic page containing:

Input waveform and output waveform on the same amplitude scale.
Band architecture: dark rectangles show the pass regions, pale rectangles show Hann transition regions, and red marks show requested band centers.
Panning architecture: upward bars show left coefficients and downward bars show right coefficients for each band.
Gain profile: linear band gains across the frequency bank.
Summary: tuning mode, bandwidth, resonator model, decay, gain profile, Dry/Wet, wet/output peaks, level action, duration, source-channel behavior, and random seed.

The plot describes the requested center frequencies. For tuned integer delays, the exact delay-derived frequency and cents error are reported numerically in the Info window.

Practical notes

Narrow bands isolate smaller spectral regions; wider bands pass more source material into each resonator.
High decay produces longer resonance. In Finite echoes mode, the maximum number of taps is explicit; in Single feedback comb mode the decay continues recursively through the available Sound duration and optional tail.
tail_ms is useful when a source ends immediately after an excitation and you want to retain the subsequent ringing.
Random panning is the only nondeterministic part of the processor when random_seed = 0. Use a non-zero seed to reproduce the same spatial assignment.
Because Praat’s Hann band-pass is a frequency-domain, zero-phase filter, its impulse response is symmetric in time. Very abrupt material can therefore show pre/post ringing associated with the band extraction itself.