One sound, not five layers
Spectral Gesture Forge is a native-Praat spectral resynthesis instrument for creating a single short, expressive sound-design gesture. It draws one recording from each active source folder, extracts acoustic material and combines the recordings inside one evolving spectral synthesis process.
Listen to examples
These three existing examples illustrate three different gesture characters using the same random seed. They are musical demonstrations, not a controlled comparison between the new reverberation modes.
Impact
Rapid spectral onset, concentrated body and decisive release.
SGF_Impact_seed1
Bloom
Early low-frequency body followed by a wider, resonant spectral expansion.
SGF_Bloom_seed1
Reverse
An accelerating accumulation toward a late peak with a sharp release.
SGF_Reverse_seed1
The timbres also depend on the chosen recordings, envelope settings and spectral-fusion strength.
Getting started in Praat
- Run Spectral Gesture Forge from AudioTools → Spectral.
- Choose Folders to browse (1–5). Leave Folder paths empty to select folders through Praat, or supply their full paths separated by semicolons (
;). - Choose Random or Spectral complementarity for source selection.
- Set the Duration (3–7 seconds), select a Gesture character, and adjust envelope sharpness, spectral variation, fusion strength and stereo movement.
- Under Reverberation, leave None for the original dry fusion, or select Per-source, Global or Per-source + global. Set amount, average decay and tail mode. Enable More reverb settings only if you want to adjust diversity, pre-delay or damping in a second dialog.
- Leave Keep result in Praat enabled and click OK. The script renders offline; it can play the result and draw a diagnostic figure.
- To save the sound, select the new SGF_… Sound in Praat’s Objects window and use Save → Save as WAV file…. There is no automatic export option in the current interface.
Controls at a glance
| Control | What it does | Default |
|---|---|---|
| Folders to browse (1–5) | Number of source folders to select when the paths field is blank. | 1 |
| Folder paths | Optional semicolon-separated folder paths; bypasses folder-selection dialogs. | Blank |
| Source selection | Random selection or spectral-complementarity search across candidates. | Random |
| Duration (s) | Length of the principal gesture, limited to 3–7 seconds. An extended reverberation tail can make the rendered Sound longer. | 4.5 |
| Gesture character | Impact, Bloom, Reverse, Pulse or Hybrid. | Impact |
| Envelope sharpness (%) | Controls ADSR attack/decay timing and curvature. | 75 |
| Spectral variation (%) | Controls differences among source-band envelopes. | 60 |
| Spectral fusion strength (%) | Changes the power-mean magnitude-combination rule. | 85 |
| Motion preset | Sweep, Converge, Expand or Oscillate. | Expand |
| Stereo width (%) | Degree of separation between spectral regions across the stereo image. | 70 |
| Motion amount (%) | Strength of stereo trajectory and relative motion. | 65 |
| Reverb mode | None, Per-source, Global, or Per-source + global. | None |
| Reverb amount (%) | Strength of the selected reverberation pathway. | 25 |
| Average reverb decay (s) | Nominal decay time, internally bounded to 0.2–5 seconds. | 1.8 |
| Tail mode | Contained keeps reverb within the gesture; Extended allows an additional decaying tail. | Contained |
| More reverb settings | Opens a second Praat dialog for decay diversity, pre-delay and damping. | Off |
| Random seed | Reproduces selection, envelopes and spatial/reverb variations when sources and settings are unchanged. | 1 |
| Keep result in Praat | Retains the generated Sound in the Objects list so you can listen or save it. | On |
| Draw visualization | Draws a diagnostic figure in Praat Picture. | On |
| Play result | Auditions the generated Sound after rendering. | On |
Additional reverberation settings
Tick More reverb settings to open a second dialog after the main form. The defaults are Reverb decay diversity: 50%, Pre-delay: 30 ms and Reverb damping: 40%. These settings are not shown in the main form unless requested.
Five gesture characters
Impact
A quick attack, especially strong low-frequency body and short high-frequency flashes. Designed for compact, dramatic sonic events.
Bloom
Low and middle frequencies appear early; upper bands emerge later and spread outward, with a longer resonant evolution.
Reverse
Energy accumulates toward a late shared peak. A swift release follows, while a small amount of low resonance may remain.
Pulse
Two or three closely spaced internal impulses are shaped as one unified gesture, rather than a phrase or repeating rhythm.
Hybrid
Different spectral bands use different gesture behaviours, producing less predictable yet coordinated transformations.
Reverberation as part of the sound
The optional reverb engine offers two different musical relationships between sound and acoustic space: reverberation inside spectral fusion and reverberation after fusion. It is disabled by default.
None
Runs the original spectral fusion without reverberation. Use this as the reference for comparing the instrument's dry synthesis.
Per-source
Each selected recording receives its own reverberant response. The wet sound enters the fusion engine as additional spectral material, with separate, evolving reverberant envelopes. This is not simply reverb added to a completed mix.
Global
Applies stereo reverberation to the already-fused gesture, creating a common surrounding acoustic space while retaining its articulated source spectrum and stereo trajectory.
Per-source + global
Combines both methods: source-specific resonances contribute to the fused sound, followed by a lighter global stereo reverb. The global wet amount is reduced to avoid excessive buildup.
Why source-level reverb is different
Every source can have a different decay, pre-delay, damping and diffusion profile. Instead of being mixed after synthesis, its reverberant spectrum becomes an additional member of the shared power-mean fusion. Dedicated reverberant envelopes allow the tail to emerge as the dry source-band envelopes recede. The result can sound like a single, imaginary acoustic object with its own internal resonance.
How the reverberation is generated
Praat generates finite, source-specific impulse responses with early reflections, diffuse decaying energy and frequency-dependent damping, then applies them through FFT convolution. This is a finite convolution reverberator, not a recursive feedback network. Source-dependent and seeded differences give each recording its own reverberant character.
| Setting | Musical effect |
|---|---|
| Reverb amount | Controls the strength of the reverberant contribution. |
| Average decay | Sets the approximate centre of the source-dependent decay range. |
| Decay diversity | At low values source reverbs are similar; higher values create greater differences in decay and related characteristics. |
| Pre-delay | Separates the attack from the start of the reverberant response. |
| Damping | Shortens high-frequency reverberation relative to lower-frequency energy. |
| Tail mode | Contained: the tail is faded inside the selected gesture duration. Extended: the sound can continue after the gesture, with up to approximately four additional seconds as the reverberant energy decays. |
What happens inside the sound?
Source selection
One usable file per folder, chosen randomly or through a bounded comparison of spectral centroid, bandwidth, tonality and six-band energy profiles.
Acoustic material
A detected strong attack determines the source region. Short recordings can be read more slowly, within limits, before their contribution fades.
Spectral shaping
Each source contributes to six overlapping, approximately logarithmic spectral regions, controlled by independent source-band envelopes.
Phase-aware spectral fusion
At each analysis frame, the script combines the selected recordings' spectral magnitudes using a weighted power mean, with phase derived from a soft dominance-weighted complex combination. Inverse FFT and overlap-add construct the resulting signal. If per-source reverberation is enabled, each source's wet signal joins this same fusion process as an additional spectral member.
Lower fusion strength approaches weighted magnitude summation; higher strength emphasises components supported across the active sources. A material gate excludes nearly silent source-band contributions, and the engine compensates band energy to preserve envelope control.
Analysis uses a 1,024-point FFT with Hann windows and 75% overlap, and six smooth spectral regions centred between approximately 80 Hz and 12 kHz. Sources are resampled to a common 44.1 or 48 kHz working rate. Stereo sources are downmixed for analysis, and a new stereo image is synthesized.
Articulation and resonant memory
Up to 30 direct source-band ADSR envelopes control attack, decay, sustain, release, onset delay and curvature. When per-source reverb is enabled, separate reverberant envelopes allow a decaying spectral memory to emerge during the release without simply lengthening every direct ADSR.
Stereo motion
The script computes changing global and band-dependent positions. Low frequencies remain closer to the centre; higher bands can spread and move independently. Left and right receive different spectral weightings and controlled upper-band phase offsets. Optional global reverb is applied after the stereo fusion, using related but distinct stereo responses.
Sweep
Motion across the stereo field.
Converge
Spectral components move toward the centre.
Expand
A focused gesture opens into a wider stereo field.
Oscillate
Curved, time-varying stereo motion.
Output, saving and repeatability
A standard run creates one stereo Sound in Praat's Objects window, normally named SGF_Impact_seed1 or similarly for the selected character and random seed. When enabled, the script also auditions the result, displays a diagnostic Praat Picture, and prints a detailed rendering report to the Info window.
Keep result in Praat is enabled by default and should remain enabled if you want to save the generated sound after processing. Turning it off removes the Sound after playback. The source recordings themselves are never overwritten.
The Info-window report includes the selected source files and regions, rendering settings, spectral and stereo measurements, and when active, source-specific reverberation details and the final rendered duration. Randomness is seeded for repeatability, provided the source files and settings remain unchanged.
With Tail mode: Extended, the rendered Sound can exceed the specified 3–7 second gesture duration because reverberation continues after the direct event. The additional tail is bounded and faded before the Sound ends.
Practical notes
Do I need Python or a DAW?
No. The spectral fusion, source reverberation and global reverberation all run offline within native Praat.
Which recordings can I use?
The folder scanner accepts WAV, AIFF/AIF/AIFC, FLAC and MP3 file extensions. Actual decoding depends on the installed Praat version and the individual files.
Does the script use every file in each folder?
No. It selects one usable recording per active folder. The complementarity mode considers a bounded set of candidates (up to 12 analysed and 40 attempts per folder), without scanning nested subfolders.
What happens with short recordings?
The script can slow the read position after the detected attack, up to twice the original span. If the material still runs out, its dry envelopes fade. With Per-source reverb, independently generated decays may continue where permitted by the tail mode.
Which reverberation mode should I try first?
Start with Per-source to make reverberant material part of the spectral fusion. Use Global for a common acoustic space around an already-fused gesture, or Per-source + global for both.
Why doesn't the audio stop exactly at the requested duration?
With Extended tail mode and reverb enabled, the reverberation can continue after the principal gesture. Use Contained if you need the sound to end within the specified 3–7 seconds.
Where is the exported WAV file?
The script no longer exports automatically. Leave Keep result in Praat on, select the resulting Sound in the Objects window, and save it manually as a WAV file.
How can I get the same result again?
Use the same input folders, files, random seed and all processing settings. Reverb randomness is controlled separately so changing the reverb mode does not change the already chosen source regions or direct ADSR shapes.
Is the result always perceived as one fused sound?
No. The engine aims for coherent spectral fusion, but perceptual unity depends on the acoustic sources, fusion strength, envelopes and reverberation settings.
Created by Shai Cohen, Department of Music, Bar-Ilan University. Part of Praat AudioTools (MIT License).