Fractal Spectral Hologram — User Guide

An experimental, analysis-driven sound-synthesis instrument. It extracts spectral components from a recording and resynthesizes evolving, geometrically transformed families of partials, resonances and diffuse textures, entirely within Praat.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel License: MIT License Repo: Praat AudioTools
Contents:

What this does

Fractal Spectral Hologram uses a recording as the source of new sonic material. It analyzes a succession of overlapping Gaussian-window spectral frames, estimates prominent partial frequencies and amplitudes, then synthesizes new oscillators at geometrically related frequencies. The generated partials have continuously integrated phase, independent attack and decay behavior, and optionally delayed, fragmented or diffuse components.

The result may retain a perceptible relationship to the source or develop into a substantially different crystalline, metallic, resonant, or evolving spectral texture. The source recording is not simply passed through a spectral filter: the processed (wet) signal is assembled from newly synthesized components.

Native Praat: analysis, additive resynthesis, temporal envelopes, sound mixing and visualization run inside Praat. No Python, Parselmouth, external DSP program or temporary audio file is required.
Think of this as synthesis, not an effect preset library: the presets change the musical behavior of the engine. They are starting points; activating Edit_parameters exposes their underlying controls.

What does “fractal spectral hologram” mean here?

The instrument derives several frequency structures from one analyzed partial. Each copy is placed relative to a chosen frequency center, using a geometric ratio. Positive exponents expand distances from that center; negative exponents contract them. A copy can also arrive later, decay differently, or emerge progressively as the source plays.

fₚ(t) = c + z(t)ᵖ · [ f(t - Δₚ) - c ] c = Zoom_centre_Hz z(t) = Fractal_zoom → Zoom_end over the source duration p = 0, +1, +2, ... ; optionally −1, −2, ... Δₚ = |p| × Echo_delay_s (rounded to analysis frames) weight(p) = Source_level if p = 0, else Level_decay^|p|

For example, with c = 0 and z = 2, a 440 Hz analyzed partial can generate components at 880 Hz and 1760 Hz. With c = 700 Hz, the transformation instead expands or contracts each partial’s distance from 700 Hz. Mirror_levels adds copies on both sides of the original scale hierarchy.

When Zoom_end differs from Fractal_zoom, the ratio changes over time. Depth_growth can introduce the outer levels progressively, so the spectral structure unfolds rather than appearing all at once.

Fractal-inspired, not recursive: each copy is derived from the analyzed partial or captured mode, not from the previous generated level. “Hologram” is a compositional metaphor for multiple related spectral images, not a simulation of optical holography.

Quick start

  1. In Praat, select exactly one Sound (mono or stereo). If the source is an audio file, import it as a Praat Sound first.
  2. Run Fractal_Spectral_Hologram.praat.
  3. Choose a preset. Subtle Shimmer is a restrained starting point; Fractal Storm and Glass Fracture explore more radical results.
  4. Set Dry_wet in the main dialog: 0 gives a copy of the original, 1 gives the synthesized signal alone.
  5. Enable Edit_parameters to open the two detailed dialogs, or choose Custom to open them automatically. Otherwise, the selected preset runs with its predefined parameters.
  6. Use Draw_visualization to show the source, wet spectrogram and component map. Disable it when concentrating on the audio or reducing total processing time. Play_result controls automatic playback.
  7. The result is created as a new Sound named <source>_hologram_<preset>; the original Sound remains unchanged.
Offline processing: the script can take longer to compute than the recording’s duration, particularly for long sounds and dense presets. This does not mean that the operation has failed.

Three ways to obtain the source partials

1. Tracked partials

Detects prominent peaks in successive spectral frames and links nearby frequencies across time. Each track drives resynthesized partials that follow the input’s changing pitch and energy. When a track ends, its envelope may continue to ring.

Used by: Subtle Shimmer, Crystal Echo, Fractal Storm and Glass Fracture.

2. Struck modes

Captures a fixed set of prominent spectral modes. Subsequent rises in energy around each mode’s frequency excite a bank of decaying resonances. It can turn attacks into metallic, inharmonic ringing.

Used by: Metallic Bell.

3. Frozen modes

Captures a fixed set of spectral frequencies and sustains them, with optional slow modulation, fractal expansion and decay. With automatic capture, the script searches for the loudest region.

Used by: Holographic Freeze.

Capture details: Capture_region_s = 0 uses the whole input for the captured mode set. A positive region width selects a segment centered at Capture_time_s; setting the capture time to 0 selects an automatic intensity maximum. For tracked partials, the capture controls do not apply.

For irregular or noisy material: frozen mode capture can fall back to weaker local maxima. If struck mode capture finds no usable prominent peaks, the script substitutes a harmonic mode set based on 220 Hz; it reports the fallback in Praat’s Info window.

Presets

PresetSource modeMain synthesis behaviorSelected settings
Subtle ShimmerTrackedOctave-related partial halo with a small diffuse layer; retains a relatively direct relationship to the source.Center 0 Hz; zoom 2; 2 positive levels; diffuse 0.15; ring T60 1.2 s.
Crystal EchoTrackedDelayed, geometrically scaled copies that persist as separate resonances.Center 400 Hz; zoom 1.5; 4 levels; 170 ms delay per level; ring T60 2 s.
Fractal StormTrackedExpanding and contracting spectral cluster; outer levels enter progressively, with substantial diffuse energy.Center 700 Hz; zoom 1.30 → 1.75; ±4 levels; depth growth 1; diffuse 0.45.
Holographic FreezeFrozenA captured modal spectrum becomes a sustained, slowly evolving texture.Auto-capture 0.4 s; zoom 1.25 → 1.40; 4 levels; 1500 ms attack; 2.5 s tail.
Metallic BellStruckBand-onset-triggered resonator bank with an inharmonic geometric spectrum and long decay.Center 90 Hz; zoom 1.618; 3 levels; ring T60 8 s; diffuse 0.
Glass FractureTrackedBright high-ratio partials broken into short, deterministic fragments.Center 150 Hz; zoom 2.3; 3 levels; fragments 11 Hz at density 0.38; diffuse 0.
CustomUser-selectedOpens the full parameter dialogs for designing a new combination of analysis, mapping, envelopes and texture.Starts from the base control values shown below.

Presets assign multiple interacting parameters, including partial count, source contribution, attack, decay, and analysis settings. The values above summarize their musical identity; use Edit_parameters to view or change the full set.

Controls

Main dialog

ControlDefaultMeaning
PresetSubtle ShimmerChoose one of the six musical presets or Custom.
Dry_wet0.8Linear mix: 0 = unchanged source copy, 1 = synthesized-only output.
Edit_parametersOffOpen both detailed dialogs after applying the chosen preset.
Draw_visualizationOnDraw source/wet spectrograms and a generated-partial map.
Play_resultOnAutomatically play the newly created Sound.

Analysis and fractal map

The following are Custom starting values. Selecting a named preset overrides a subset of them.

ControlCustom defaultMeaning
Source_modeTracked partialsChoose tracked partials, struck resonator modes, or frozen captured modes.
Window_ms40Gaussian spectrogram analysis-window parameter, 10–200 ms. Larger values favor frequency resolution over temporal precision.
Hop_ms10Time step between analysis frames, at least 2 ms and no greater than Window_ms.
Partials16Maximum selected spectral peaks per analysis frame or captured modal set (1–40).
Sharpen0.5Emphasizes spectral peaks relative to a local ≈250 Hz spectral envelope; emphasis is bounded.
Min_track_frames3Minimum length of a tracked partial before it is retained; short tracks are discarded.
Fractal_zoom1.5Starting geometric mapping ratio (0.25–4).
Zoom_end1.5Ending geometric ratio; a different value creates continuous zoom evolution.
Zoom_centre_Hz600 HzStationary frequency axis of the mapping, between 0 Hz and Nyquist.
Fractal_levels3Number of nonzero positive copy exponents (0–8).
Level_decay0.6Copy-level amplitude weighting: decay raised to |p| (0–1.5).
Mirror_levelsOffAdd negative exponents (contracting copies) alongside positive ones.
Source_level0.5Weight of the p = 0 resynthesized source-level partial, not a dry-audio mix.
Depth_growth0Progressive fade-in of successively deeper copy levels; 0 makes all levels available immediately.

Time, texture and capture

ControlCustom defaultMeaning
Attack_ms20Attack smoothing for tracked and frozen partial envelopes; 0 is immediate.
Ring_s1.0 sReference T60 decay time at 1 kHz: the time for a 60 dB fall in amplitude. 0 removes the ring persistence.
Ring_tilt0.3Frequency dependence of decay; lower partials can ring longer, higher ones shorter.
Echo_delay_s0Additional delay per fractal exponent magnitude, rounded to the analysis grid (0–2 s).
Fragment_rate_Hz0Rate of deterministic per-component fragment patterns; 0 disables fragmentation.
Fragment_density0.4Proportion of the available fragment events that occur (0–1); relevant when fragment rate is positive.
Diffuse_mix0Balance between coherent sinusoidal lines and noise-modulated spectral textures; 0 is coherent only.
Blur_Hz40 HzApproximate width of the filtered-noise modulation in the diffuse path (2–2000 Hz); not global FFT-bin blur.
Tail_s1.0 sExtra output duration following the original sound, for decays and delayed copies (0–30 s).
Capture_time_s0Capture center time for struck/frozen modes; 0 requests automatic selection when a finite region is used.
Capture_region_s0.5 sTime-span used to measure the captured mode set; 0 means the whole source.
Two different ways to hear the source: Source_level changes the p = 0 synthesized partials inside the wet engine. Dry_wet mixes the original recording with the completed wet output. Set Dry_wet = 1 to evaluate the synthesis on its own.

How the processing works

  1. Time-resolved analysis. Extract each channel and compute a Gaussian-window Praat Spectrogram, with overlapping analysis frames. The analysis band stops at the lower of 12 kHz and 95% of Nyquist.
  2. Spectral structure. Estimate a local frequency-dependent spectral envelope; select significant peaks, with log-parabolic refinement of their frequencies and amplitudes and bounded prominence-based sharpening. A reference sine provides runtime amplitude calibration.
  3. Partial organization. Either track peaks between frames, capture a modal spectrum triggered by band-energy rises, or capture a sustained frozen mode set. Discard short tracks and allow accepted tracks to ring when their excitation stops.
  4. Fractal copy controls. For each source partial or captured mode, build multiple mapped-frequency and amplitude-envelope controls. Copy levels can have independent delays, weights, frequency-dependent decay, progressive depth and time-varying zoom.
  5. Additive resynthesis. Create a newly synthesized wet signal by summing sinusoids with continuously integrated phase. The engine skips some inactive or inaudible contributions where possible, without intentionally changing the generated partial structure.
  6. Optional texture. Add deterministic fragmentation and a separately controllable diffuse path based on filtered noise modulation. With Diffuse_mix = 0, this diffuse contribution is disabled.
  7. Output assembly. Match the wet signal’s RMS to the source over the source duration, fade its last few milliseconds, then apply Dry/Wet mixing. If the final peak exceeds 0.99, attenuate the completed mix for safety.
Not a phase vocoder: the spectrogram is used for analysis, but the signal is not reconstructed by inverse FFT or overlap-add of processed spectral frames. It is generated as an oscillator-based additive synthesis from analyzed partial trajectories and modal data.

Channels, duration and level

Dry/Wet = 0: the result is sample-identical to the original, including its duration, although the script still performs its internal analysis and synthesis before creating that copy. For Dry_wet = 1, the wet path contains synthesized components only, subject to wet level matching and the safety gain.

Visualization

When Draw_visualization is enabled, Praat draws three coordinated panels for channel 1, all referenced to the output time axis:

The legend reports exponent and weight for every level. The information strip summarizes analysis settings, decay and delay controls, diffuse settings, wet gain, dry/wet, safety gain, output length, the number of drawn component segments and audio processing time. The frequency range adapts to the generated partials, up to the display limit.

Performance note: spectrograms and thousands of plotted line segments can require considerable extra time. The reported Processing time is measured before visualization is drawn. Disable Draw_visualization for timing comparisons.

Processing time and practical considerations

Fractal Spectral Hologram is an offline synthesis instrument, not a real-time plug-in. Every detected or captured partial can drive several separately weighted oscillator layers, potentially for the full source duration and additional ringing tail. Long inputs, high sample rates, many partials, dynamic zoom, mirrored levels and the diffuse path can increase calculation time substantially.

The Info window reports analysis frames, tracking or captured-mode statistics, synthesized region count and oscillator-seconds. It also distinguishes all potentially active oscillator-seconds from the portion for which sine terms are actually evaluated. This helps explain computational cost without confusing it with the number of analyzed partials.

Efficiency without simplifying the sound: the engine avoids computing some copy levels while they are outside the audible frequency range or have not yet entered through Depth_growth. Phase integration is retained as needed for continuity. Reported oscillator-seconds represent computational work estimates, not playback duration.

For a demanding preset such as Fractal Storm, first try a shorter excerpt. To reduce processing cost while designing a texture, lower Partials or Fractal_levels, avoid unnecessary mirrored levels, and disable visualization. These changes may alter the sound, so preserve a preferred setting before adjusting it.

Historical and compositional context

The modern language of fractal geometry is associated especially with Benoît Mandelbrot, who introduced the term in the 1970s and studied forms organized by self-similarity at different scales. In composition, self-similarity can also be a structural principle without requiring a mathematically exact fractal.

Here the concept is applied to spectral and temporal composition. An analyzed partial becomes a family of frequency-related partials; the family can expand, contract, arrive in staggered time, decay independently, or combine into a diffuse texture. The musical outcome comes from the interaction between the source’s spectral behavior and the chosen synthesis controls, rather than from repeating an entire prerecorded sound.

The term hologram is metaphorical: it refers to multiple related projections of spectral material, not an optical model. For historical background on fractal geometry, see IBM — Fractal geometry.