Ligeti Micropolyphonic Choir Machine

Transform a single recording into a dense field of related voices. Build a relatively stable choir with detuning and staggered entrances, or sculpt evolving relationships between voices at the millisecond scale.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel Script version: 2.1.1 (2026) License: MIT Praat AudioTools repository

What the script does

Ligeti Micropolyphonic Choir Machine creates a many-voice texture from one selected Praat Sound. The voices can be offset, detuned, faded and distributed across the stereo field. Two processing engines offer different ways of organizing this sonic mass.

Legacy: a cloud of whole voices

Each voice is a transformed copy of the complete recording, with its own fixed time offset, nominal detuning, duration variation and envelope. This is the original voice-cloud behavior.

Microtemporal: voices with changing internal timing

Each voice reads the source through a displacement that can evolve during the sound. Voices may move between temporal coherence and dispersion, with optional independent source-reading scatter and compensation for coherence-related level changes.

These are two different sound-making approaches, not simply two quality settings. The first emphasizes the accumulation of staggered and detuned lines; the second focuses on changes in the temporal relationships within the mass.

What this does not claim: an increase in microtemporal dispersion does not automatically mean a greater perceived acoustic volume. Depending on the source and settings, the result may sound broader, rougher, more diffuse or more strongly phase-colored. Listening remains essential.

Quick start

  1. Open or create a sound in Praat, and select exactly one Sound object.
  2. Run Ligeti_Micropolyphonic_Choir_Machine.praat.
  3. Choose a preset. For a familiar layered-choir sound, start with Static Spectral Fog. For a gradual change in the internal texture, try Emergence or Convergence.
  4. Choose Speed mode. Balanced is the default; Full Quality retains the original sampling rate for processing.
  5. Optionally enable Microtemporal settings to open a second dialog and edit the evolving-offset engine. The dialog appears when the active mode is Microtemporal, including presets that select it automatically.
  6. Choose Stereo spread, Wet/Dry, and Output level; then run. The processed Sound appears in Praat alongside the original.
Preset priority: Experiments A–D and the four new musical presets automatically select the Microtemporal engine even if the first dialog shows Legacy. They also override specific controls, described below. The five original presets use the processing mode you choose.

Presets: what to expect

The preset menu contains Custom, five original voice-cloud presets, four controlled experiments, and four newer musical behaviors. Each run can produce a different arrangement unless you specify a repeatable random seed.

Original voice-cloud presets

These preserve the established combinations of pitch distribution, whole-voice offsets, duration variation and fading. They are intended primarily for the Legacy processing mode.

PresetVoicesOffsetsNominal pitchDuration variationMusical character
Static Spectral Fog60±0.5 s±8 cents±2%Concentrated, soft-edged detuning with more voices near the center of the pitch and offset distributions.
Fracturing Mass60±1.0 sup to ±35 cents±8%Detuning range increases with voice index, building a mixed population of relatively stable and more divergent layers.
Stereo Torsion50±0.6 sup to ±25 cents±5%Permitted pitch deviation grows toward the right side of the stereo field; stereo spread is always enabled.
Bimodal Web40±0.8 s−20 to −5 / +5 to +20 cents±10%Alternating lower and upper detuning groups create two interwoven pitch regions.
Breathing Field40±1.2 s±12 cents±8%Broad, uniformly distributed cloud. The original preset name does not imply a cyclic breathing envelope.

Original preset gain/fade values: Fog 0.8 / 50 ms; Fracture 0.7 / 30 ms; Torsion 0.9 / 20 ms; Bimodal 0.85 / 15 ms; Breathing 0.8 / 40 ms. The table shows nominal pitch-control ranges, not the full pitch change after varispeed duration variation. Except for Stereo Torsion, Stereo_spread remains under user control.

Pitch and duration are linked in these presets. The script changes the sample-rate interpretation of a voice, so Duration_variation also alters its actual pitch. For example, ±10% duration variation can introduce substantially more pitch change than a nominal ±20-cent control. The Info window now reports both declared and effective pitch deviation.

Controlled experiments: isolate the timing

Experiments A–D are listening and comparison conditions, rather than finished musical textures. They all use 24 voices, a mono result, 100% wet, no pitch or duration variation, and Match source RMS output level. The source's own RMS may not be fully attainable if the peak safety limit intervenes.

PresetTemporal behaviorUse it to hear…
A · SynchronousAll voices aligned; micro-offset 0; macro-offset 0.A coherent reference for the same voice count.
B · Static Micro OffsetFixed, independently assigned micro-offsets within ±10 ms.The effect of constant sub-10-ms alignment differences.
C · Evolving Micro OffsetSlow stochastic micro-offset trajectories within a nominal ±10 ms range.Time-varying coherence, contrasted with B's fixed offsets.
D · Macro CloudWhole-voice offsets within ±500 ms; no micro-offset evolution.Broad temporal staggering rather than microscale variation.

The experimental conditions do not enable read scatter or coherence compensation by default. Use the same input and random seed where appropriate when comparing runs. Listening tests should begin with the unprocessed source, then A, B, C and D.

Musical Microtemporal presets

The following presets work with changing temporal relationships. All four disable nominal pitch and duration variation, use zero macro-offset, apply 100% coherence compensation and use a rate of one cycle over the output timeline. The initial scatter mode is Free; it may introduce audible roughness, especially with pitched sounds.

PresetVoicesMicro-offsetRead scatterBehavior and likely musical use
Emergence40±20 ms±40 msMoves from a more coherent opening into a dispersed, internally active mass. Try for the gradual emergence of texture from a focused sound.
Convergence40±20 ms±40 msThe reverse trajectory: temporal independence recedes toward a synchronized ending. Useful for contraction or resolution.
Breathing Coherence40±8 ms±25 msOne expansion/contraction cycle of microtemporal dispersion over the output duration.
Coherence Waves48±8 ms±20 msThree voice groups expand and contract with phase-shifted envelopes, creating successive waves of activity.

These presets leave the first dialog's stereo choice, Wet/Dry blend and output-level selection available unless otherwise stated. Their time-trajectory depth can be reduced by the Doppler limiter when the requested movement is too fast for the material's duration. Details are reported in Praat's Info window.

Controls and settings

First dialog: general controls

ControlDefaultWhat it changes
PresetCustomSelects a stored voice distribution, an experiment or a musical evolution.
Processing_modeLegacyChooses fixed whole-voice processing or evolving microtemporal processing. Experiments and new musical presets override this selection.
Microtemporal_settingsOffOpens a second dialog with micro-offset, evolution, scatter, compensation, reconstruction and timeline controls whenever Microtemporal is active.
Speed_modeBalancedFull Quality processes at the source rate. Balanced works at up to 22,050 Hz; Fast at up to 11,025 Hz. Lower-rate options save processing time at the expense of bandwidth.
Number_of_voices60Number of layers in Custom; rounded to an integer and never below two.
Time_offset_range_s0.8 sWhole-voice temporal offset range for Custom. The meaning of negative offsets depends on the processing engine.
Duration_variation0.05Random varispeed factor around 1; also affects pitch. A value of 0 disables this variation.
Max_pitch_cents15Nominal pitch-deviation range. A value of 0 disables nominal detuning.
Random_seed_(0_=_new_each_run)00 generates a new seed for each run; a positive integer reproduces the random choices on repeat runs with matching settings.
Stereo_spreadOnPlaces mono-derived voices across a stereo field using equal-power panning. When off, channels are preserved where applicable.
Attack_fade_ms30 msTime-domain edge fade; preset-specific in most preset modes.
Voice_gain1.0Base gain before the per-voice 1 / sqrt(N) factor.
Wet_dry_percent80%Blend between the processed choir and the original recording. Experiments force 100% wet.
Output_levelPeak normalizePeak normalization to 0.95, source-RMS matching with clipping protection, or no output-level adjustment.
Draw_visualization / Play_resultOnOptional graphs and automatic playback after processing.

Second dialog: Microtemporal settings

The advanced dialog is shown only if Microtemporal_settings is enabled and the active processing mode is Microtemporal. It starts with the chosen preset's values, which you can then adjust before rendering.

ControlInitial Custom valueMeaning and permitted range
Micro_offset_ms20 msMaximum evolving or static micro-displacement per voice, 0–100 ms.
EvolutionStochastic driftNone, Static, Emergence, Convergence, Breathing, Waves or Stochastic drift.
Evolution_rate_Hz0.2Rate in Hz (0.01–5), or number of cycles across the timeline (0.1–50), depending on Rate_unit.
Rate_unitHzChoose absolute frequency or duration-relative cycles. The field label retains “Hz” even when it represents a cycle count.
Source_divergence_ms0 msIndependent per-voice source read advance between 0 and the chosen limit, up to 2000 ms. Different from delaying a voice's onset.
Read_scatter_ms0 msAdditional independently randomized read-position changes, 0–200 ms, scaled by the evolution envelope.
Scatter_modeFreeFree permits abrupt per-grain changes outside the Doppler limiter; Smooth incorporates scatter into the constrained trajectory.
Coherence_compensation_pct0%0–100% dynamic correction of level changes associated with coherent summing or cancellation.
Grain_ms40 msReconstruction-window duration, 10–200 ms, with 50% overlap and complementary windows.
Doppler_limit_cents15 centsRequested limit for pitch change caused by smooth delay evolution; 0 disables the limit. Sample resolution can constrain precision.
TimelinePreserve source durationKeep the source's duration and fade late material, or extend the output for delayed voices.
A practical distinction: Smooth scatter follows the Doppler-constrained trajectory. Free scatter can generate abrupt read-position jumps that are not covered by the reported smooth-trajectory Doppler figure, and can produce pitch jitter or roughness on tonal recordings.

How the engines work

Legacy: how detuning and duration interact

The original engine computes a pitch ratio and a duration factor, then performs one combined sample-rate conversion:

pitch_ratio = 2 ^ (nominal_cents / 1200)
duration_factor = 1 + random_duration_variation
combined_factor = pitch_ratio / duration_factor

This approach is efficient, but it is varispeed, not independent pitch shifting and time stretching. The actual pitch difference depends on both nominal detuning and duration variation. The script reports the effective deviation for the generated voices.

For negative legacy offsets, the beginning of the copy is trimmed and the remaining audio starts at the output's time zero. Positive offsets delay playback. The legacy output buffer extends beyond the working source by the nominal offset range plus 0.5 seconds.

Microtemporal: two interleaved overlap-add streams

The microtemporal engine reconstructs each voice using two streams of overlapping sin²-windowed segments. Each segment reads from a fixed integer-sample displacement; the displacement can change across successive segments. Static offsets therefore do not require per-sample pitch manipulation, while changing offsets may create a Doppler-like change in pitch.

voice_displacement(t) = macro_offset
                      + micro_trajectory(t)
                      - source_read_advance
                      + optional_read_scatter(t)

Macro offset shifts the voice relative to the whole composition. Micro trajectory changes relative timing within the sound. Source divergence advances the material read by a voice without representing a new onset. Read scatter varies reading positions at the reconstruction-grain level.

When the Doppler limit makes the requested trajectory impossible, the script attempts to reduce its depth so that the specified start and end conditions remain feasible. The Info window reports the reduction, constrained trajectory deviation, or any remaining endpoint warning.

Coherence compensation is not a loudness model

Nearly synchronized voices can sum strongly; differently shifted copies can also cancel. Optional coherence compensation estimates independent-voice energy from the actual source-reading positions in approximately 50 ms frames and applies a smoothed, bounded gain correction.

At 100%, it aims to reduce the unintended changes in level caused by coherence. It does not guarantee constant perceptual loudness or prove an increase in perceived acoustic volume. The degree of correction, smoothing and limiting all affect the result.

Output, reproducibility and visualization

Choosing an output level

OptionBehaviorWhen to use it
Peak normalize to 0.95Scales a non-silent output so its peak reaches 0.95; can amplify quiet material as well as attenuate loud material.Everyday rendering and preset exploration.
Match source RMSMeasures the source's total-power RMS and the output RMS over the corresponding source-length span, then applies a safe gain capped by a 0.99 peak ceiling.Comparisons in which gross level differences would otherwise distract. Praat reports if the requested match was not reached.
NoneLeaves the rendered level unchanged and reports peaks above 1 that may clip on export.Controlled post-processing and external level management.

Output naming: sourceName_ligeti_PresetName. The original Sound remains available. For reproducible randomized arrangements, enter a positive Random_seed and record the chosen settings. A seed of zero generates a new seed, which the script prints in the Info window.

Duration: the default Microtemporal timeline preserves the source duration; the alternative extends the result to include late voices. Legacy retains the longer output-buffer behavior. Sample rate: Balanced and Fast may downsample during processing and then restore the original sample rate, but lost high-frequency detail is not restored by upsampling.

What appears in the graphs

Legacy visualization

Source and processed waveforms, spectrogram, voice cloud by nominal pitch and whole-voice offset, stereo placement by nominal pitch, and a summary of the render.

Microtemporal visualization

Source and processed waveforms, spectrogram, trajectories for the first ten voices, temporal-dispersion display with all-voice range and mean ± standard deviation, and processing summary.

The Microtemporal panels are derived from the displacement data used in the rendering, not from an illustrative preset curve. The Info report also includes per-voice displacement snapshots, effective pitch deviation, scatter details, level-compensation behavior and any trajectory constraints.

Listening and composition tips

Recommended first listening sequence: Experiment A → Experiment B → Experiment C → Emergence → Breathing Coherence. This separates the question “what does a millisecond-scale timing difference do?” from the musical question “what happens when those relationships evolve?”

Historical and compositional context

György Ligeti's micropolyphony brings together numerous closely interwoven vocal or instrumental lines so that intricate internal activity may be perceived as a slowly transforming sonic mass. Works often associated with this approach include Apparitions, Atmosphères, the Requiem, Lux aeterna and Lontano.

The Choir Machine is inspired by the perceptual result of such dense polyphony. It is not a reconstruction of Ligeti's specific contrapuntal techniques or scores. The Legacy engine layers transformed recordings; the Microtemporal engine adds continuously changing timing relationships. Neither claims to generate independently notated counterpoint.

The newer engine also opens a different line of exploration associated with microsound and temporal coherence: rather than composing only the succession of sonic events, a composer can shape the relationships between components inside a sound. The audible outcome depends on the source, reconstruction settings and distribution of the voices.

Further reading