Praat AudioTools / Synthesis · Spectral · Composition

Undertone Field

Build evolving polyphonic fields below the source pitch through sample-rate reinterpretation, mathematical undertone anchors and independent playback-rate voices.

Author: Shai CohenVersion: 4.0 (2026)Engine: Sample-rate reinterpretation / constant-rate varispeedOutput: Stereo Sound

What this does

Undertone Field creates lower-pitched, slower-moving versions of a selected Sound. Unlike an ordinary transposition effect, it organizes these versions around a descending series of mathematical undertones. In Pure mode, each undertone is one voice. In Deviated mode, each undertone becomes a group of closely related voices that gradually drift apart in time.

Each voice plays at its own constant speed below 1.0. Pitch and duration therefore change together: a rate of 0.5 sounds one octave lower and plays for twice as long; a rate of 0.2 sounds substantially lower and lasts five times as long. The changing relationships between the voices arise from their different playback speeds, not from a continuously fluctuating pitch envelope.

Mathematical backbone

Integer, odd, even, custom and geometric undertone series provide the pitch anchors.

Independent voices

Controlled deviations create polyphony, microtonal beating and increasingly separated events.

Compositional duration

Hear the entire slowed field or impose a fixed duration with an explicit ending fade.

Undertone anchors with deviated companion voicesThree groups below the source, each with an anchor and two slightly deviated constant-pitch voices extending through time. Source1/21/31/4StartTime →
Conceptual illustration, not a measurement of a preset: each thicker line is an undertone anchor; thinner lines are close deviations. Longer lines represent lower playback rates.
The governing constraint: all synthesized voices remain below the source pitch. This is an undertone composition engine, not an unrestricted upward/downward random-varispeed processor.

Quick start

  1. Select exactly one Sound in Praat's Objects window (minimum duration 50 ms).
  2. Open Undertone_Field.praat and run it.
  3. Choose a preset. For the clearest demonstration of the new architecture, try Deviated Harmonic Field, Microtonal Undertone Cloud or Expanding Undertone Field.
  4. Choose a Monitor setting. Undertones only reveals the new material without the original masking it.
  5. Choose Natural duration to hear complete voices, or Fixed length to set a particular composition length.
  6. Optional: enable Show advanced settings to adjust series stretch, filters, spatial layout, entry spacing and the source window.
  7. Run the script. It creates a stereo Sound named <source>_undertones_<presetName>, prints a detailed Info report and, if requested, draws a Picture-window visualization and plays the result.
Useful starting point: Preset = Deviated Harmonic Field, Monitor = Undertones only, Duration = Natural. The preset sets up four anchors with three voices each (12 voices total).

The undertone series

Each anchor has a nominal playback-rate factor r = 1 / denominator. For the integer series:

r(n) = 1 / (n + 1)(1 + stretch)   for n = 1, 2, 3, ...

With stretch = 0, the rates are 1/2, 1/3, 1/4, 1/5.... Increasing stretch spreads successive undertones farther downward; negative stretch compresses the spacing.

Series modeAnchor ruleMusical character
Integerdenom(n) = (n+1)^(1+stretch)Standard descending undertones: 1/2, 1/3, 1/4...
Odd denominatorsdenom(n) = (2n+1)^(1+stretch)At zero stretch: 1/3, 1/5, 1/7...
Even denominatorsdenom(n) = (2n)^(1+stretch)At zero stretch: 1/2, 1/4, 1/6...
Custom listUser-entered denominators greater than 1Direct control of the anchors; stretch is not applied. Missing entries continue upward in steps of +1.
Geometricdenom(n) = 2 × [1.25^(1+stretch)]^(n-1)At zero stretch: 2, 2.5, 3.125, 3.906...

Rates are computed without rounding to an integer override sample rate. A minimum reinterpretation clock of 100 Hz is enforced, and the resulting rates must remain strictly below 1. The Info report identifies anchors whose requested rate had to be raised to this floor.

Pure and Deviated undertones

Pure undertonesDeviated undertones
VoicesOne voice per anchorOne exact anchor voice plus companions per anchor
PitchMathematical rate onlyRates around each anchor, with deviations specified in cents
TimeDifferent constant rates create a slow-layer fieldNearby constant rates cause additional beating and temporal divergence
Maximum8 voices8 undertones × 8 voices = 64 voices

How deviations are selected

Voice 1 of each group is the exact anchor. Additional voices receive stratified randomized deviations, nominally within ±Deviation_cents. A gap around zero discourages companions from being nearly identical to the anchor. The available range is reduced automatically when required to limit upward companion deviations toward a ceiling 50 cents below the source, respect the 100 Hz reinterpretation-clock floor, and (when enabled) keep neighboring undertone groups from overlapping. A custom anchor already closer than 50 cents to the source is not moved downward by that ceiling; the strict requirement is that every voice remain below the source.

Random_seed = 0 draws a new seed each run. A positive seed reproduces the same generated deviations under the same settings. The seed actually used appears in the Info report. Every voice retains its individual playback rate throughout its life; deviations do not fluctuate continuously.

Audible distinction: small deviations make close beating and gradual drift; wider deviations can create more conspicuous rhythmic separation. Because playback rates remain below 1, no companion rises above the original source.

Presets

The script provides 11 presets plus Custom: seven earlier series-based treatments and four new polyphonic constructions. The table gives the actual default configurations defined in v4.0. Levels are nominal dB values relative to the source RMS, before any output-wide peak-safety attenuation.

PresetStructureColour / spaceMonitor defaultDry / field level
Pure Harmonic4 × 1; integer; linear curveFixed LP 3000 Hz; arcOriginal + Undertones0 / −3 dB
Stretched Cloud5 × 1; integer +0.20; GaussianFixed LP 2600 Hz; arcOriginal + Undertones0 / −2 dB
Compressed Cluster4 × 1; integer −0.20; exponentialFixed LP 3200 Hz; arcOriginal + Undertones0 / −4 dB
Bass Spectral6 × 1; integer; inverse curveFixed LP 2200 Hz; arcOriginal + Undertones−3 / 0 dB
Spectral Haze5 × 1; geometric +0.12; GaussianFixed LP 3600 Hz; arcOriginal + Undertones0 / −4 dB
Dark Undertow4 × 1; integer −0.10; exponentialFixed LP 1600 Hz; arcOriginal + Undertones−4 / +1 dB
Microtonal Dense7 × 1; geometric +0.05; flat curveFixed LP 3000 Hz; arcOriginal + Undertones0 / −5 dB
Deviated Harmonic Field new4 × 3; integer; ±30 centsAdaptive; balanced clustersUndertones only0 / 0 dB
Deep Polyphonic Undertones new4 × 2; odd series; ±20 cents; inverseWarm; balanced clustersOriginal + Undertones−12 / 0 dB
Microtonal Undertone Cloud new2 × 7; integer; ±15 cents; flatNatural; balanced clustersUndertones only0 / 0 dB
Expanding Undertone Field new5 × 3; integer; ±12 cents; group entries every 60% of source-window durationAdaptive; balanced clustersUndertones only0 / 0 dB
Preset behavior: choosing a named preset overrides its synthesis structure, series, curve, levels, spectral mode and spatial settings. It does not prevent you from explicitly choosing a different Monitor, Duration, or Random seed. For named presets, Dry_level_dB in the main form is superseded by that preset's dry level; use Custom when you want to set dry and field levels independently. Advanced settings open with the active preset's values and can then be edited.

Controls and Advanced settings

Main form

ControlFunction
PresetChoose Custom or one of the 11 named configurations.
SynthesisPure: one voice per undertone. Deviated: several voices around each undertone.
Number_of_undertones1–8 mathematical pitch anchors.
Series_mode, Gain_curveSelect the pitch series and relative level profile across anchor groups.
Voices_per_undertone1–8 voices per group, used in Deviated mode.
Deviation_centsNominal companion deviation range, 0–600 cents; may be restricted for safety or group ordering.
Random_seed0 = fresh random field per run; a positive integer = repeatable field.
MonitorPreset default, Undertones only, Original + Undertones, or Original only.
Dry_level_dB, Undertone_level_dBSeparate levels for the source and the synthesized field. Named presets set their own dry level.
Duration, Fixed_duration_sComplete natural duration or intentional fixed-length output with fade.
Spectral_balanceNatural, Warm, Dark, Adaptive, or Fixed low-pass.
Show_advanced_settingsOpen two additional configuration dialogs.
Draw_visualization, Play_resultControl visual feedback and automatic audition.

Advanced dialog 1: series and level

Stretch_factor (−0.5 to +0.5), Custom_denominators, Rolloff_dB (0–24), Anchor_emphasis_dB (−12 to +12), Keep_groups_in_order, Normalise_layers, Low_register_comp_dB (0–12 dB/octave) and Lowpass_Hz for Fixed low-pass mode.

Advanced dialog 2: space and time

Spatial_layout, Stereo_spread_pct (0–100), Cluster_width_pct (0–100), Micro_delay_ms (0–20 ms), Entry_spacing_pct (0–400% of the source-window duration), Fade_s, Source_start_s, Source_end_s and Max_output_s (1–1800 s).

The Advanced dialogs appear after the main form. Changes made there apply to the current run. The base main-form settings and the selected preset are not dynamically rewritten on-screen.

Monitoring and level balance

Version 4 separates the original sound from the undertone bed. Instead of relying on a stack of loosely related per-layer gains, it balances individual voices and then measures and levels the complete undertone field once.

MonitorAudible result
Preset defaultThe monitoring choice specified by the preset (or Original + Undertones for Custom).
Undertones onlyThe generated field alone; ideal for judging the new polyphonic texture.
Original + UndertonesThe original source window, at its configured level, mixed with the undertone field.
Original onlyThe dry source alone, useful as a reference. The script still constructs and analyzes the undertone field.

Relative gain curves

Flat treats all groups equally. Linear decreases by the selected rolloff per successive group. Exponential uses a faster initial fall with a saturating profile. Gaussian emphasizes the middle undertones. Inverse increases toward deeper undertones before normalization of the group levels. The script normalizes the curve relative to its maximum before applying the overall field target.

What the level controls mean

Dry_level_dB sets the dry contribution relative to the source-window reference. Undertone_level_dB targets the measured total RMS of the stereo undertone bed relative to that same reference, over a defined level-measurement span. For simultaneous entries this span starts at the beginning of the window; for staggered group entries it starts when the last group enters. The reported full-output RMS may differ from the target.

Optional per-voice RMS matching is limited to ±18 dB. Low_register_comp_dB is off by default; when enabled it adds relative compensation per octave below the source and may change the slope of the chosen curve. The report flags curve reversals and level-matching limits.

Finally, the script checks the completed selected output. If its peak exceeds 0.99, one attenuation-only gain is applied to the whole output. This protects against clipping, but may lower both dry and processed components together; the Info report states the attenuation and identifies whether the dry, wet or combined signal triggered it.

Important for listening comparisons: the nominal dry/field values are not guarantees of final perceptual loudness. Low-frequency weighting, overlap between voices, source content and final peak safety all matter. Review the measured RMS, peaks, and field-to-dry relationship in the Info window.

Duration, entries and source windows

Natural duration

Every voice runs through its slowed source window. Total output reaches the last voice end plus a 50 ms release pad.

Fixed length

The composition ends at Fixed_duration_s. Voices not completed by then are intentionally cut; an optional raised-cosine fade closes the output.

The natural duration of a voice is source-window duration / playback rate, plus its entry offset. For example, a 12-second source window played at 1/5 lasts approximately 60 seconds from that voice's entry. A companion slightly below 1/5 lasts a little longer. In natural mode the script checks the complete resulting duration against Max_output_s (default 300 s), rather than silently shortening the composition.

Source_start_s and Source_end_s in Advanced settings select a portion of the original Sound without copying a library of files. The default 0 / 0 means the whole Sound; an end of 0 means the source's actual end. A 5 ms edge fade is applied at an explicitly cut source-window boundary.

Entry_spacing_pct offsets successive undertone groups by a percentage of the source-window duration. For example, 60% makes each group begin 0.6 window durations after the preceding group. Micro_delay_ms introduces small additional offsets between voices. These change the temporal relationships but not their constant playback rates.

Memory and long renders: up to 64 extended voices can be expensive to render. The script estimates peak working memory and stops if a Natural render exceeds Max_output_s; it does not automatically switch to Fixed length. Use a shorter source window or explicit fixed duration for exploratory work.

Spectral balance

All voices are generated by slowing the source, which already lowers its spectral bandwidth. Version 4 offers different filtering choices rather than imposing one fixed low-pass treatment on every result.

ModeBehavior
NaturalNo extra low-pass shaping; retains the bandwidth available after transposition, subject to the subsonic guard.
WarmGentle low-pass centered at 4500 Hz (2000 Hz smoothing).
DarkStronger low-pass at 1800 Hz (800 Hz smoothing).
AdaptiveFor voices with transposed bandwidth above 2500 Hz, cutoff follows sqrt(2500 × voice bandwidth); otherwise no additional low-pass.
Fixed low-passUses Lowpass_Hz from Advanced settings, with a 100 Hz smoothing width; reproduces the older fixed-cutoff approach.

The script also attempts a 30 Hz subsonic high-pass guard on each voice. For efficiency, these filters are implemented on the source-window material before sample-rate reinterpretation, at frequencies scaled by the planned voice rate. A low-pass that would lie at or above that voice's available transposed bandwidth is omitted.

Extreme slowdown can push most musical energy below the audible range. More aggressive RMS normalization is not necessarily an improvement: it can amplify low-frequency energy or noise. The subsonic guard and reported gain adjustments help identify this behavior.

Stereo organization

The generated undertones are distributed across a stereo field with equal-power left/right coefficients. Advanced settings offer two layouts:

Arc

Anchor groups progress across the stereo panorama from left to right, with the overall width governed by Stereo_spread_pct.

Balanced clusters

Deeper groups occupy more central positions. Other groups alternate across the field using estimated level and duration to reduce uneven left/right energy. Companions spread around their group's position according to their deviation.

Cluster_width_pct sets the maximum lateral offset of companions around their anchor group. Every voice is clipped to the available stereo pan range. The layout does not automate pan position over time.

For mono input, the dry component is placed in the center at equal power. For stereo input, the dry component preserves the original left and right channels; the undertone engine itself uses a mono working mix. If L+R nearly cancels (its RMS is less than one-quarter of the two-channel RMS reference), the undertone engine switches to the stronger source channel and reports this choice. Inputs with more than two channels are folded to mono for dry and processed material.

Check the reported left/right balance, channel correlation and mono fold-down, especially with wider clusters or strongly decorrelated source material.

Processing architecture

  1. Prepare a source window. Preserve the original Sound; build a mono working source for the undertones while retaining the original stereo channels for the dry path.
  2. Generate anchors and voices. Derive the mathematical undertone rates; in Deviated mode, assign bounded companion deviations with a reproducible random seed.
  3. Plan the composition. Determine start offsets, natural voice lengths, required output duration and any deliberate fixed-length cuts.
  4. Filter each source-domain voice. Apply low-frequency guarding and the selected spectral balance using rate-adjusted filter frequencies. Optionally RMS-match the voice within ±18 dB.
  5. Reinterpret and resample. Override the sampling frequency to source sample rate × voice rate, then resample at the original rate (50-point sinc). Pitch and duration remain altered.
  6. Accumulate the undertone bed. Apply group gain, anchor/companion weighting and equal-power pan, adding each voice at its scheduled start time.
  7. Level the completed bed once. Measure the selected level span and set the undertone-field level relative to source RMS.
  8. Assemble the selected monitor output. Add the separately controlled original source if required; apply the Fixed-length ending fade if selected.
  9. Apply peak safety. Attenuate the completed output only if its absolute peak exceeds 0.99, and document the gain change.

There is no PSOLA pitch preservation in this path: slower playback always means lower pitch and greater duration. The processing is offline inside Praat; no Python, Parselmouth or external executable is required.

Picture visualization and diagnostic report

When Draw_visualization is enabled, the Picture window displays a multi-panel overview of the actual output and its planned voice structure:

PanelReading it
Source windowThe selected original material shown on the output time axis as an amplitude reference.
Output L / RSeparate output waveforms, making level and stereo differences visible.
Voice mapOne horizontal line per voice: vertical position = cents below source; starting point = entry; length = sounding duration. Thicker lines are anchors; marked ends indicate truncation.
Spectral balanceLong-term source, undertone and (when mixed) output spectra versus logarithmic frequency, with reference markers for transposed spectral centroids.
Group levelsAchieved group levels versus requested relative levels, with source, dry and field references.
Stereo fieldEach voice placed by stereo position and measured level; contextual mono fold-down information.
SummaryChosen mode, duration, voice count, peak safety and key level measurements.

The Info window provides two principal tables. Undertone anchors reports denominator, effective rate, cents, curve/compensation, requested and achieved group balance, permitted deviation range, entry and pan. Voices reports each anchor/companion's rate, reciprocal rate, pitch, deviation, natural length, start, rendered length, percentage cut, filter, RMS normalization, final relative level and pan.

Further sections give duration and render time, dry/undertone/combined RMS and peaks, interference or reinforcement of the voices, any global peak-safety attenuation, and stereo diagnostics. The report also identifies clamped rates, restricted deviations and target levels that could not be reached.

Picture panels and Info values are analysis and monitoring aids; they do not imply real-time synthesis or continuous automatic evolution of playback rates. The changing texture comes from simultaneous voices proceeding at different fixed rates.

Output, limits and listening advice

Propertyv4.0 behavior
Created object<source>_undertones_<presetName>
FormatStereo Praat Sound at the source sample rate
TimeNatural: final voice end + 0.05 s. Fixed: exactly the selected duration, with optional ending fade.
Pitch rangeAll synthesized voice rates are strictly between 0 and 1. Upward companion deviations are bounded by a -50-cent ceiling where possible.
Source materialOriginal selected Sound remains available; the script works on an extracted source window.
Working limits1–8 undertone groups; up to 8 voices each; configurable output-duration ceiling and memory estimate.
Peak protectionOutput gain reduced only when absolute peak exceeds 0.99; applied adjustment is reported.

Suggested listening experiments

  1. Use the same sustained or melodic source with Pure Harmonic and Deviated Harmonic Field; listen for beating, independent motion and increasing temporal separation.
  2. Compare Microtonal Undertone Cloud in Natural mode with a Fixed-length version. Short rendering windows can conceal the later development of slow voices.
  3. Try Expanding Undertone Field with Undertones only; inspect the voice map to follow the staggered entrances.
  4. Switch a Custom patch between Natural, Warm and Adaptive spectral balance to assess how filtering affects clarity without changing the anchor mathematics.
  5. Check the Info report for output-wide peak attenuation before concluding that the undertone field is too quiet; lowering the dry contribution may preserve more of the intended wet level.
Practical limitation: very low playback rates produce long files and can shift much of the sound below 30 Hz. If a Natural render exceeds its duration ceiling, the script stops and asks for an adjustment rather than silently trimming the result. An exact audition should be verified in the installed version of Praat.