Partial Editing Resynthesis — Sinusoidal Texture Resynthesis — User Guide

Frame-based additive resynthesis that selects strong spectral partials, edits their frequency and amplitude, and reconstructs a new mono texture.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel Version: 0.7 (2026) License: MIT License Repo: https://github.com/ShaiCohen-ops/Praat-plugin_AudioTools

Contents:

What this does Sinusoidal analysis & resynthesis Quick start Presets Controls Processing pipeline Channels, randomness & level Visualization Historical & compositional context

What this does

Partial Editing Resynthesis analyzes short overlapping frames, finds a limited number of strong spectral peaks, edits their frequencies and amplitudes, and rebuilds the sound as a sum of sine waves. The script's form is titled Sinusoidal Texture Resynthesis because the result is intentionally a new additive texture rather than a transparent reconstruction of the source.

For each frame, the processor searches between Min_frequency and Max_frequency, repeatedly takes the strongest remaining spectral peak, suppresses a ±40 Hz region around that peak, and continues until Max_partials_per_frame has been reached or no sufficiently strong peak remains.

What is a partial? A partial is one sinusoidal component of a complex sound. Harmonic sounds often contain partials near integer multiples of a fundamental frequency, but this processor does not assume harmonicity: it simply selects strong short-time spectral peaks.

Sinusoidal analysis and resynthesis

A sinusoidal model represents sound as a collection of time-varying sine components. Here the analysis is deliberately simpler and more textural than a full sinusoidal-tracking model:

Peaks are detected independently in each analysis frame.
Peak frequencies are kept on FFT-bin centres before user editing.
There is no trajectory tracking that follows one partial from frame to frame.
The source phase is not copied into the new oscillators; resynthesis uses sine generators evaluated on absolute time.
There is no separate stochastic/noise residual model.

Overlapping Hann-windowed grains are summed and then divided by the exact accumulated window envelope. This gives stable overlap-add reconstruction even when the user chooses a window/hop ratio that is not a textbook constant-overlap-add setting.

Quick start

Select exactly one Sound. Mono and multichannel sources are accepted, but the output is always mono.
Run Partial_Editing_Resynthesis.praat.
Start with Clean Texture Resynth or Dense Partials.
Use Max_partials_per_frame to move between sparse and rich reconstructions.
Use Transpose_semitones for musical transposition and Additional_frequency_scale for an additional multiplicative spectral scale.
Choose stable jitter for continuity or frame-random jitter for the older grainier texture.

Presets

Preset	Key overrides
Custom	No override.
Clean Texture Resynth	Freq jitter 0.5 Hz; amp jitter 0.02; 20 partials.
Diffuse Texture	8 Hz; 0.3; 15 partials.
Sparse Partials	5 partials; 2 Hz; 0.1.
Dense Partials	30 partials; 1 Hz; 0.05.
Pitch Up Octave	+12 st; 1 Hz; 0.05.
Pitch Down Octave	−12 st; 1 Hz; 0.05.
Spectral Scale Up	Additional frequency scale 1.5; 2 Hz; 0.1.
Spectral Scale Down	Additional frequency scale 0.7; 2 Hz; 0.1.
Glassy Shimmer	15 Hz; 0.4; 20 partials; max frequency 12 kHz (clipped to Nyquist if necessary).
Robotic	No frequency or amplitude jitter; 12 partials.
Whisper Ghost	4 partials; 10 Hz; 0.5; max frequency 6 kHz.

Controls

Control	Default	Meaning
Window_length	0.060 s	Duration of each Hann analysis/resynthesis frame.
Hop_size	0.015 s	Time between frame starts. The final frame is shifted when necessary so the source end is covered.
Min_frequency / Max_frequency	60 / 8000 Hz	Peak-search band. Max is clipped to source Nyquist.
Max_partials_per_frame	15	Maximum number of strongest peaks resynthesized in each frame.
Freq_jitter	3 Hz	Maximum detuning amount around each detected FFT-bin frequency; negative values are clamped to 0.
Amp_jitter	0.1	Relative amplitude variation; internally clamped to 0…1.
Jitter_behaviour	Stable per-frequency	Stable mode deterministically maps each FFT bin to a fixed detune/gain offset. Frame-random mode generates new random offsets in each frame.
Transpose_semitones	0	Frequency multiplier `2^(st/12)`.
Additional_frequency_scale	1.0	Additional positive multiplier applied to every resynthesized partial frequency.

The final partial frequency is approximately (detected_frequency + jitter) × 2^(transpose/12) × additional_scale. Components below 20 Hz or at/above working Nyquist are omitted from the synthesized grain.

Processing pipeline

For multichannel input, measure each source channel and choose the highest-RMS channel as the analysis driver. This avoids cancellation that could occur if anti-phase channels were averaged.
Create a mono output buffer and a separate overlap-add window-sum buffer at the original sample rate and duration.
For every frame: extract a Hann-windowed segment, compute its Spectrum and Ltas, find strong peaks one by one, and suppress ±40 Hz around each chosen peak before searching again.
Apply the selected frequency/amplitude jitter, transposition and additional frequency scaling.
Synthesize all accepted partials for that frame as sine waves under an analytic Hann window.
Overlap-add every grain into the output.
Divide by the exact accumulated Hann-window envelope.
If the result is non-silent, apply Scale intensity: 70.

Channels, duration, randomness and level

Output: always mono.
Multichannel analysis: highest-RMS source channel only; the other source channels are not mixed into the synthesis.
Sample rate: source sample rate.
Duration: source duration.
Stable jitter: deterministic for a given FFT-bin identity and parameter set; no random seed is required.
Frame-random jitter: uses unseeded random values, so repeated runs can differ.
Final level: non-silent output is set to 70 dB SPL in Praat's intensity convention. This is not 70 dBFS and is not peak normalization.

Output name: <source>_resynth_<preset>.

Visualization

The figure shows original and resynthesized waveforms on one shared amplitude scale, original/output spectrograms, and a summary of partial count, jitter, transposition and frequency scale.

Multichannel display detail: analysis uses the highest-RMS channel, but the “Original Spectrogram” visualization currently displays source channel 1. On multichannel material these can therefore represent different source channels.

Historical, technological and compositional context

Sinusoidal analysis/resynthesis became a major digital sound-modeling technique in the 1980s. McAulay and Quatieri's 1986 sinusoidal representation estimated short-time sinusoidal frequencies, amplitudes and phases from spectral peaks and tracked those components through time. In computer music, Serra and Smith's 1990 Spectral Modeling Synthesis extended spectral modeling by separating a sound into deterministic sinusoidal trajectories and a stochastic residual.

This Praat tool deliberately takes a more reduced, compositional approach. It keeps only a limited set of strong peaks in each frame, does not track partial trajectories across frames, and does not resynthesize a noise residual. Those omissions are not merely technical limitations: they are what make sparse, robotic, diffuse and ghost-like textures possible. The source is treated as material for a new additive reconstruction rather than as something that must be reproduced transparently.