IRCAM SuperVP Transform — User Guide

Praat AudioTools bridge to IRCAM SuperVP for analysis-driven spectral transformation, resynthesis and creative processing. Praat supplies F0 and Intensity analysis; the Python backend builds verified SuperVP commands and control curves, runs the engine, and returns both the sound and a manifest of the process.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel Version: 3.2.4 (2026) License: MIT License Repo: GitHub
Contents

What this does

SuperVP Transform connects Praat's analysis tools with the IRCAM SuperVP command-line engine. It is designed as an offline analysis–resynthesis environment rather than a real-time effect.

  1. Praat exports the selected Sound object(s) to WAV.
  2. Praat analyses pitch and intensity. Pitch can be estimated automatically with a two-pass range finder or with a user-selected range.
  3. The Python backend converts analysis into SuperVP BPF/control files and assembles the exact command line.
  4. SuperVP performs the transformation/resynthesis.
  5. Praat imports the result and can draw the analysis, control curves, spectrograms and executed commands.
Current scope: 16 processing modes plus Diagnostics. Newer SuperVP features are capability-checked against the help text of the installed build; unsupported features are not silently substituted with a different process.

Quick start

  1. Install and activate a licensed copy of IRCAM SuperVP. On the current Windows setup the executable is typically C:\Users\User\SuperVP\bin\supervp.exe.
  2. Place supervp_transform.py in plugin_AudioTools/py/.
  3. Select one Sound in Praat. Select two Sounds only for Spectral cross synthesis.
  4. Run IRCAM_SuperVP_Transform.praat.
  5. Choose a processing mode, quality profile and pitch range.
  6. Click Next, adjust the mode-specific parameters, then click Apply.
  7. The dialog remains open so parameters can be adjusted and re-applied without restarting the script.
Good starting point: use Standard (v2 behaviour) when testing a mode for the first time. Use General or Voice when you want SuperVP's newer adaptive analysis options.
SuperVP is not included with Praat AudioTools. The script is a bridge to the separately licensed IRCAM engine.

Processing modes

ModeMain SuperVP mechanismWhat it does
Voice age-ageBuilt-in voice ageing/rejuvenation. Positive values sound older; negative values younger.
Gender-female / -male, optional -transenvBuilt-in gender transformation, optionally reinforced with an additional formant/envelope shift.
Age + genderTwo SuperVP passesRuns Age first, then Gender.
Flatten pitch-transke BPFMoves the measured F0 contour toward its median or a user target. Amount 1 = flat; 0 = unchanged; negative values exaggerate the contour.
Vibrato-transke BPFDeterministic sinusoidal pitch modulation, optionally applied only to voiced frames with onset fade-in.
Organic vibrato-trmdlpSuperVP randomly redraws vibrato rate and depth within user-defined ranges.
Harmoniser-transke + waveform mixingCreates a transposed copy and mixes it with the original at a fixed musical interval.
Formant shift-transenvSpectral-envelope shift with Constant, Ramp, Follow F0, Follow Intensity and Wander shapes.
Time stretch-DConstant or time-varying dilation. Shapes include constant, voiced-only, unvoiced-only and ramp.
Tremolo-ggain, -transenv, -trmdle, or -gtremoloFour tremolo engines: amplitude, spectral-envelope, constant SuperVP envelope tremolo, and native SuperVP tremolo with sine/triangle/square/saw waveforms. Depth can follow Praat Intensity.
Breathiness-GcrossCreates a whispered copy by combining the source spectral envelope with noise fine structure, level-matches it, then blends it with the original. It no longer relies on -Gnoise.
Spectral cross synthesis-GcrossCombines amplitude and frequency structure from two Sounds. Simple mode is spectral cross-coupling, not a linear source mix; Advanced mode exposes X/x/Y/y weights.
Dynamic gate-ggain BPFTime-domain output gain derived from Intensity. Useful for gating, but it does not remove noise inside the spectrum.
Spectral denoise-avseg/-OM1 → -FsubTwo-pass spectral subtraction. SuperVP first learns a noise spectrum from a noise-only time segment, then subtracts that model from the full signal.
Spectral phase diffusion-rnd_force, -rnd_voi, -rnd_stval, -rnd_ff, optional -shape 1/-rnd_sinUses SuperVP phase randomization as a creative effect. Separate control is available for non-sinusoidal material, noise above the VUF and—when sinusoidal separation is enabled—sinusoidal components.
Frequency shift-FfshiftTrue frequency shifting in Hz. Every partial moves by the same Hz amount, producing inharmonic spectra unlike pitch transposition in cents.
DiagnosticsSuperVP help interrogationReads the installed build's help and reports quality flags and creative features that are actually supported.

Quality profiles & pitch analysis

The quality profile controls analysis-window behaviour, overlap, transient handling and selected SuperVP quality options. Flags are emitted only when the installed SuperVP help confirms them.

ProfilePurpose
Standard (v2 behaviour)Compatibility profile. Keeps the established processing path with no new v3 quality flags.
GeneralGeneral-purpose high-quality analysis. With voiced F0 it uses Praat's pitch track for an F0-adaptive window (-F0 + -Mauto).
VoiceMonophonic speech/singing profile. Adds voice-oriented windowing and, where applicable, adaptive True Envelope and perceived-loudness preservation (-envpl 1).
PercussiveShorter window and transient-oriented analysis.
SmearedLong analysis window with transient preservation intentionally reduced for a more blurred/smeared result.

Adaptive True Envelope

In suitable envelope-based modes, the General and Voice profiles can use Praat's measured F0 directly in SuperVP:

-F0 f0.bpf -Afft +<95th-percentile-F0>Hz

This lets the envelope order adapt to the source rather than using one fixed envelope order.

Pitch range

The main dialog offers:

Pipeline

1. Export — selected Praat Sound(s) → temporary WAV.
2. Praat analysis — F0 and Intensity → BPF files.
3. Python bridge — analysis-derived BPFs + capability-checked SuperVP command.
4. SuperVP — spectral analysis / transformation / resynthesis.
5. Manifest — exact commands, quality flags, notes and control curves.
6. Praat import — processed WAV + optional visualisation.

BPF control files

Most control curves are two-column text files containing time value. Depending on the mode, values can represent F0 in Hz, transposition in cents, linear gain, frequency shift in Hz, stretch factor, tremolo depth, or another SuperVP control parameter.

Spectral denoise is a special two-pass process

Pass 1: -avseg start,end -OM1 noise_profile_M1.sdif
Pass 2: -Fsub noise_profile_M1.sdif -avbeta ... -avsfac ... -avrelax ... -avgamma ... -avdet ...

The first pass creates an SDIF noise model containing the mean and standard deviation of each spectral bin. The second pass performs Bark-smoothed spectral subtraction.

Important mode parameters

ModeKey parametersDefault / meaning
Flatten pitchTarget pitch, Flatten amountTarget 0 = median F0; amount 1 = flat, 0 = unchanged.
VibratoRate, Depth, Voiced only, Onset fadeDeterministic F0 modulation in cents.
Organic vibratoRate min/max, Depth min/max, New values everyDefaults: 4.5–6.5 Hz, depth 20–60, new values every 0.4 s.
Formant shiftShape, Shift, End, Modulation depth, Wander rate, SeedDepth meaning depends on selected shape.
Time stretchStretch factor, Stretch shapeDefault factor 1.5.
TremoloType, Native waveform, Intensity mapping, Rate, Amplitude depth, Envelope depth, SmoothingNative -gtremolo uses SuperVP's internal waveform; the Rate field does not control that native mode.
BreathinessAmount, SeedDefault amount 0.4.
Spectral cross synthesisControl type, Spectral cross coupling, X/x/Y/y, Source-2 fitSimple coupling 0.5 is not a 50/50 waveform mix; it cross-couples amplitude and frequency structure.
Dynamic gateThreshold below peak, SmoothingDefault threshold 30 dB, smoothing 40 ms.
Spectral denoiseNoise sample start/end, Maximum reduction, Noise estimate std factor, Relaxation, Subtraction type, Deterministic noiseDefaults: 0–0.25 s, 20 dB, 1.0, 30 ms, amplitude subtraction, deterministic handling off.
Spectral phase diffusionSinusoidal randomization, Non-sinusoidal randomization, Noise randomization at VUF, Full randomization frequency, Separate sinusoidsDefaults: 0, 0.15, 0.30, 18 kHz, separation off. Levels are fractions of π.
Frequency shiftShape, Shift, Ramp end, Modulation depth, Wander rate, SeedDefault shift −50 Hz; true Hz shift, not transposition.
Spectral denoise: choose a genuinely noise-only segment. If the segment contains a cello attack, voice, resonance or other desired source material, that material becomes part of the learned noise model and may be attenuated in the full signal.
Phase diffusion: -rnd_sin is meaningful only when SuperVP separates sinusoidal components. Enable Separate sinusoids (-shape 1) for a monophonic/single-source signal when you want independent sinusoidal randomization.

Visualization & manifest

When Draw visualisation is enabled, the Praat picture is driven by the manifest written by the Python backend. It can include:

The manifest also stores the full executed command lines (cmd_full) so the exact processing path can be inspected or reproduced.

Dry run: use Dry run (curves + commands only) to inspect analysis and generated commands without rendering audio.

FAQ / troubleshooting

“supervp.exe not found”

Point the first dialog's SuperVP executable field to the installed command-line binary. The path is remembered between sessions.

SuperVP asks for activation

Activate the installed build with the valid IRCAM activation key file before using the Praat bridge. Activation is handled by SuperVP itself, not by this script.

Spectral denoise removes part of the instrument/voice

Move the noise-learning interval to a section that contains only background noise. The learned SDIF model is used as the spectrum to subtract from the entire sound.

Phase diffusion's sinusoidal control seems to do nothing

Enable Separate sinusoids (-shape 1; mono source). Without sinusoidal separation, the script intentionally ignores the sinusoidal randomization value and still applies the non-sinusoidal/noise controls.

Dynamic gate vs. Spectral denoise

Dynamic gate changes output gain over time using -ggain. Spectral denoise estimates a noise spectrum and removes it inside the time-frequency representation using -Fsub. They solve different problems.

Pitch transposition vs. Frequency shift

-transke changes frequencies proportionally (musical transposition). -Ffshift adds the same number of Hz to every partial, usually making harmonic sounds increasingly inharmonic.

Diagnostics

Run Diagnostics after changing SuperVP versions. The backend reads the help of the installed build and reports which newer flags and creative features are actually available.