Creative Convolution
Convolve any two sounds, from natural room responses to experimental sonic hybrids. Choose traditional convolution or a moving window that scans through the kernel to create continuously changing spectral and temporal textures.
What this tool does
Creative Convolution uses a selected Source Sound and a second Sound as a convolution Kernel. The kernel can be a room impulse response, a voice, a metal resonance, a sound of water, an instrumental note or a synthetic texture. The original source remains untouched. Convolution transfers a combination of the kernel's temporal and spectral characteristics to the source.
Two engines serve different compositional purposes. Static Convolution applies one prepared kernel to the whole source. Moving-Window Convolution selects short, changing excerpts of that kernel as the source unfolds. Both use Praat's native convolution operation and render offline without Python or external DSP libraries.
Quick start
- Select exactly two Sound objects in Praat and run
Creative_Convolution.praat. - Choose one of the nine presets (or Custom), then select Static convolution or Moving-window convolution from Engine.
- Set Wet dry (%) yourself. It is independent of the preset: 0% is dry-only, 100% is wet-only, and the default is 100%.
- For Moving-window, start with Kernel window = 50 ms, Forward sweep over the whole source, Hop = 0 (auto) and Energy compensation = Off.
- In the next dialog, identify which selected Sound is the Source (dry); the other becomes the Kernel.
- For more kernel, spatial and gain settings, turn on Show parameters or select Custom.
- Render, listen, and optionally inspect the Praat Picture display. A moving-window result includes a scan map of the actual kernel positions used.
Two processing engines
Static convolution
A single prepared kernel is convolved with the complete source. This is the original Creative Convolution engine, with kernel preparation, optional finite feedback, and the full selection of spatial modes.
Best for room-like responses, static spectral imprinting, reverse convolution, and feedback treatments.
Moving-window convolution
Short excerpts are taken from the prepared kernel at changing positions. Source frames are convolved with their corresponding kernel windows, and all convolved frames are accumulated with their complete tails.
Convolve(SourceFrame[i], KernelWindow[i])
Best for evolving hybrid textures, changing resonances, and time-dependent relationships between source and kernel.
How the moving engine reconstructs sound
Source frames are two hops long and use periodic Hann windows with 50% overlap. The source windows form a partition of unity, so the contributions reconstruct the original source weighting. Each weighted frame is convolved with its own short kernel and placed at its original time, including its convolution tail. The accumulated result is not divided by an arbitrary overlap count.
A stationary kernel window serves as a reference: with the same prepared short kernel, the segmented result follows ordinary convolution up to numerical rounding. Kernel-window cuts receive a short raised-cosine edge taper only where the cut lies inside the prepared kernel, retaining the kernel's own start or end where applicable.
Moving-window controls
| Control | Default | What it changes |
|---|---|---|
| Kernel window (ms) | 50 | Length of each kernel excerpt, constrained to 10–500 ms. If the prepared kernel is shorter, the whole available kernel is used. |
| Window movement | Forward sweep over the whole source | How the kernel excerpt moves: static, constant-speed scan, ping-pong, random walk, or a sweep whose traversal follows the complete source duration. |
| At kernel end | Stop at the end | End behaviour for the ordinary Forward and Reverse scans: Wrap, Stop, or Turn back. The dedicated Ping-pong, Random walk and whole-source sweep modes use their own movement rules. |
| Scan rate | 1.0 | Kernel seconds traversed per source second for ordinary scans; 0 produces a fixed position. Whole-source sweeps calculate their traversal speed automatically from source duration. |
| Window start (%) | 0 | Initial position as a percentage of the available kernel-start range. Reverse trajectories are measured from the opposite end. |
| Hop (ms, 0 = auto) | 0 | Time between kernel updates. Auto sets half the kernel-window duration, bounded to 5–25 ms. Manual values: 2–250 ms. |
| Energy compensation | Off | Optional local gain correction: Off, Moderate or Full, referenced to kernel-window RMS values. Compensation is limited to +12 dB, and near-silent regions are not boosted. |
| Random seed | 0 | Controls Random walk reproducibility. A positive seed repeats the trajectory; 0 chooses a fresh seed. |
Window movement modes
| Mode | Behaviour | Musical use |
|---|---|---|
| Static position | A fixed local kernel excerpt throughout the source. | Reference condition or stable, selective filtering. |
| Forward scan | Moves through the kernel at the selected Scan rate and uses At kernel end when a boundary is reached. | Repeated or bounded exploration of the kernel. |
| Reverse scan | Moves backwards through the kernel at the selected Scan rate; uses At kernel end. | Reverse evolution of resonance and spectral colour. |
| Ping-pong | Changes direction at the kernel boundaries without a position jump. | Oscillating movement through the kernel. |
| Random walk | Smoothed, bounded stochastic movement; the trajectory reflects at kernel boundaries. | Organic, less predictable spectral changes. |
| Forward sweep over the whole source | Traverses from the selected forward start position to the last possible kernel position once, spanning the source duration. | One continuous compositional progression. |
| Reverse sweep over the whole source | Traverses in the opposite direction toward the kernel beginning once, spanning the source duration. | One continuous reversed progression. |
Window energy
Different kernel regions may contain very different energies. Off retains those differences; Moderate and Full progressively compensate for them. The correction is bounded, and a window more than 60 dB below the reference is left alone. Energy compensation is a creative control, not a replacement for final output level matching.
Nine presets + Custom
The same nine presets are available with both engines. Each preset defines kernel preparation, spectral/spatial character and related processing choices; it does not determine the user's Wet/Dry percentage or automatically switch engines. The same preset may sound substantially different when the moving engine is selected.
Natural Convolution
A largely unchanged kernel with a short fade-out and subtly expanded wet width. A useful reference for a natural impulse response or an arbitrary sound kernel.
Long Hall
Kernel stretch to 140%, 25 ms pre-delay and an expanded spatial field. Its Early Center / Late Wide setting is available in Static mode; Moving-window substitutes Wide Wet.
Reverse Bloom
Kernel reversal, 20 ms fade-in, 40 ms pre-delay and Wide Wet, designed for swelling convolution responses.
Spectral Imprint
Relatively little kernel alteration and short fades, suitable for imposing a voice, instrument or textural kernel onto a source.
Resonant Kernel
Uses the first 40% of the kernel, stretches it to 120%, and applies controlled-width resonance.
FDN Space
Designed for an FDN-rendered kernel, with a 20 ms pre-delay, a long 200 ms fade-out and Wide Wet.
Frozen Metal
Strong 300% resample-like kernel stretching with extended fades and an expanded wet field.
Short Texture
Extracts the first 15% of the original kernel before further preparation and applies short fades. In Moving-window mode, the window length is separately set in milliseconds.
Experimental Reverse FDN
Reversed, 200%-stretched kernel with a wide field and finite repeated-convolution feedback in Static mode. Feedback is unavailable in Moving-window mode.
Custom
Opens the detailed dialogs so kernel, gain, spatial processing, convolution scaling, and feedback can be adjusted directly.
Other controls
Main form
| Setting | Default | Function |
|---|---|---|
| Preset | Natural Convolution | One of nine sound designs or Custom. |
| Wet dry (%) | 100 | Independent of preset, 0–100. At 0% Wet the source is returned without output normalization. |
| Normalization | Peak | None, Peak or RMS (-18 dBFS), applied to audible-wet output. |
| Target peak (dBFS) | -1.0 | Peak target and clipping-protection ceiling when applicable. |
| Engine | Static convolution | Selects the original or moving-window DSP path. |
| Show parameters | Off | Opens detailed kernel and mix/space/advanced dialogs. |
| Draw visualisation / Play result | On / On | Draws a diagnostic picture and optionally plays the result. |
Detailed kernel controls
| Parameter | Action |
|---|---|
| Kernel start / end (%) | Extracts a region of the kernel before the remaining preparation stages. |
| Kernel stretch (%) | 25–400%. Resample-like scaling: duration and spectrum change together. |
| Reverse kernel | Reverses the prepared kernel, not the dry Source. |
| Kernel fade in / out (ms) | Raised-cosine fades; automatically shortened if their total exceeds kernel duration. |
| Pre delay (ms) | 0–500 ms, applied to the wet path only. |
Mix, space and advanced controls
| Parameter | Action |
|---|---|
| Wet dry (%) | Can also be adjusted in the detailed mix dialog; the final choice remains user-controlled. |
| Mix law | Linear or Equal power. |
| Match wet RMS to dry / Wet gain | Optional RMS match, then wet gain from -30 to +18 dB. Automatic match boost is capped at +60 dB. |
| Spatial mode | Mono Wet, Stereo Wet, Wide Wet, or Early Center / Late Wide (Static only). |
| Wet width / Wet pan | 0–200% Mid/Side width; -100 to +100 constant-power pan. |
| Convolution scaling | Sum, Integral, Normalize, or Peak 0.99. Moving-window substitutes Sum for Normalize / Peak 0.99. |
| Outside domain | Zero or Similar. Moving-window uses Zero. |
| Feedback / iterations | 0–90%, 1–4 finite iterations; only in Static mode. |
| Protect against clipping | Optional attenuation if the final output exceeds its selected safety ceiling. |
Kernel preparation
Spatial-kernel preparation → Pre-delay and convolution (Static or Moving-window) → Wet level / mixing
The moving-window engine scans the prepared kernel, not the unedited original. Trimming, stretching, reversing and fading therefore change what the window explores.
Kernel stretch is not pitch-preserving: 50% produces approximately half the duration and an octave-higher spectral scale, whereas 200% doubles the duration and shifts the spectrum down approximately one octave. The kernel is resampled to the source sample rate when necessary.
In Moving-window mode, internal window cuts are softened with a raised-cosine taper of up to 10% of window duration (maximum 5 ms), while natural boundaries of the complete prepared kernel are retained. The scan does not change the playback speed within a kernel window.
Spatial processing
The output is stereo. With a mono Kernel, Stereo Wet and Wide Wet create a complementary stereo pair using slow, opposing amplitude modulation of the kernel tail. A stereo Kernel is convolved channel-wise; an additional Mid/Side width control acts on the resulting wet field. Mono Wet uses a mono source and kernel for the convolution and places the result into the stereo output.
kR(t) = k(t) × [1 − d × m(t)]
The first 5 ms of the mono kernel remain centred; the modulation is faded in over the following 15 ms. Stereo Wet uses modulation depth 0.5, and Wide Wet uses 0.9. The complementary-sum property applies to a mono Source; it does not generally apply to channel-wise convolution of a stereo Source.
Early Center / Late Wide combines a centred early response with a later wide field in Static mode. The Moving-window engine instead uses Wide Wet and reports this substitution.
Dry/Wet, normalization and output length
Dry/Wet is independent of the preset. Every preset starts with the percentage you choose in the main dialog (or revise in the detailed mix dialog). The default is 100% Wet. Existing preset sound designs still affect the kernel and wet processing, but their former mix percentages are not applied.
Equal power: dry gain = cos(mix × π/2); wet gain = sin(mix × π/2)
0% Wet is an exact dry-only bypass at the sample level for the original channels, with the original source duration and no final normalization. The output object is still stereo: a mono source is duplicated into left and right channels. At nonzero Wet, the complete convolution tail is retained and the existing short end fade is applied only when a wet tail extends beyond the source.
For Static convolution, tail duration follows the prepared full kernel. For Moving-window convolution, the wet support follows the source plus the selected short kernel-window duration and any pre-delay; the final frame tail is retained. The source waveform itself is not overwritten.
When enabled, Wet RMS matching adjusts the active wet region relative to the source, with a +60 dB boost cap and silence protection. Final output normalization can be None, Peak or RMS (-18 dBFS); peak protection is attenuation-only when needed. These operations do not provide perceptually exact loudness matching for A/B comparisons.
Visualisation
The Praat Picture display retains the suite's waveform-and-summary design. With the Static engine, it shows the Source, prepared Kernel, wet convolution, Output and processing summary.
With Moving-window enabled, an additional Kernel scan map plots source time horizontally and position within the kernel vertically. The coloured bands show the actual local kernel windows used for each convolution frame; a line follows their start positions. The prepared-kernel panel also indicates the region visited by the scan.
The summary and Info window report the selected engine, window length, movement, scan speed, hop, frame count, energy compensation, output level, processing time and any unsupported-setting substitutions.
Creative workflows
Conventional convolution reverb
Select an impulse response as Kernel and use Static convolution. Natural Convolution or Long Hall are good starting points.
Voice × environmental sound
Use speech or singing as Source and a changing environmental recording as Kernel. Compare a static 50 ms window with a moving 50 ms window.
Instrument × resonant object
Convolve a flute, string or percussion recording with a piano, glass or metal sound. Use a whole-source sweep for one long change in the spectral imprint.
Stochastic spectral exploration
Use Random walk with a reproducible seed to wander through a noisy or composite Kernel. Compare Energy compensation Off and Moderate.
Spatial resonance
Use Stereo Wet or Wide Wet with a moving window to explore complementary left/right kernel colour, then adjust Wet Width.
Feedback textures
Use Static convolution and Experimental Reverse FDN when iterative feedback is central to the effect. Moving-window intentionally does not apply feedback.
Technical notes and limitations
- Praat: requires Praat 6.3+; no Python, plug-ins, network access or external DSP libraries are required.
- Inputs: exactly two selected Sound objects. A role-selection dialog identifies the Source; this interactive step is part of normal use.
- Channels: mono or stereo input is supported. More than two channels are folded to mono and reported. Final output is always stereo, not true multichannel convolution.
- Offline rendering: Moving-window uses multiple short convolutions, so processing time and memory depend on source length and Hop. A 20,000-frame limit prevents excessive requests.
- Moving kernel length: 10–500 ms; if the prepared kernel is shorter, its available length is used.
- Static features not reproduced by Moving-window: feedback; Early Center / Late Wide; per-frame Normalize or Peak 0.99 scaling; Similar outside-domain behaviour. The Info window documents substitutions rather than silently ignoring them.
- Reproducibility: Random seed controls the moving Random walk. The complementary stereo-pair kernel generation uses its existing deterministic modulation procedure.
- Output naming: based on the Source and preset, for example
Source_NaturalConvolutionorSource_ExperimentalReverseFDN; the engine is not added to the name. - Original files: the script creates working copies and does not modify the selected Source or Kernel.
This processor is intended for creative convolution and compositional exploration. A moving window can yield spatially suggestive or speech-like changes in certain materials, but such perceptual results depend on the Source and Kernel; they are not guaranteed by the algorithm.