Feedback-Aware Convolution — User Guide

Builds an event-driven impulse response from the source's own intensity or pitch structure, then applies that IR back to the source by convolution.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel Version: 0.5.1 (2026) License: MIT License Repo: https://github.com/ShaiCohen-ops/Praat-plugin_AudioTools

Contents:

What this does Quick start Presets Parameters Detection & IR generation Convolution & mix Visualization Output behavior

What this does

Feedback-Aware Convolution analyzes the selected Sound and turns detected acoustic events into an impulse response (IR). The analysis can follow either intensity or pitch. Each accepted event becomes a short Gaussian impulse whose amplitude is derived from the measured feature value. The source is then convolved with this event-driven IR.

The name describes a self-referential control relationship: the source determines the IR that is later applied back to that source. There is no audio feedback loop and the convolution itself is a standard fixed-IR convolution after the analysis stage has finished.

Offline process: analyze the complete source → detect events → build one IR → convolve → trim the wet result to the source duration → mix wet and dry.

Quick start

Select exactly one Sound object in Praat.
Run Feedback_Aware_Convolution.praat.
Choose one of the four built-in presets, or Custom.
For Custom, choose Intensity or Pitch, set the threshold and minimum spacing, then set impulse and Wet/Dry controls.
Click OK. If no event passes the detector, the script exits without creating an output.

Output name: <source>_feedback_<preset>.

Presets

Built-in presets override the analysis/detection fields shown below. Wet_dry_percent, Draw_visualization, and Play_result keep the values currently shown in the form.

Preset	Mode	Threshold	Minimum spacing	Impulse duration	Amplitude mapping
Custom	Form value	Form value	Form value	Form value	Form value
Sparse Intensity	Intensity	75 dB	0.08 s	0.5 ms	0.8
Dense Intensity	Intensity	60 dB	0.03 s	0.3 ms	1.2
Melodic Pitch	Pitch	150 Hz	0.05 s	0.4 ms	1.0
Rhythmic Events	Intensity	72 dB	0.04 s	0.2 ms	1.5

Melodic Pitch also sets the pitch-analysis range to 80–600 Hz.

Parameters

Control	Default	Behavior
Parameter_type	Intensity	Choose intensity-based or pitch-based event detection.
Detection_threshold	70	Interpreted as dB in Intensity mode and Hz in Pitch mode.
Minimum_spacing_s	0.05 s	Refractory interval between accepted events. While the feature remains above threshold, a new event can be accepted once this interval has elapsed.
Pitch_floor_Hz / Pitch_ceiling_Hz	80 / 600 Hz	Used only in Pitch mode. Custom bounds are internally ordered; equal bounds are separated by 1 Hz.
Impulse_duration_s	0.0003 s	Gaussian event width. Internally clamped to at least three source samples.
Amplitude_mapping	1.0	Scales the normalized event amplitude after feature-value mapping.
Wet_dry_percent	50%	Clamped internally to 0–100%.
Draw_visualization / Play_result	yes / yes	Draw the analysis summary and/or play the result.

Detection and IR generation

Intensity mode

The source is converted to Intensity with a 75 Hz minimum pitch, 1 ms time step and mean-energy subtraction enabled, then converted to an IntensityTier. The script scans its tier points in time order. An event is accepted when the value is above the threshold and the minimum spacing from the previous accepted event has elapsed.

Pitch mode

The source is analyzed with To Pitch at a fixed 1 ms time step and the selected pitch floor/ceiling. The Pitch object is sampled every 1 ms. Undefined F0 values are ignored. A defined value above the threshold can generate an event subject to the same minimum-spacing rule.

Not an edge detector: the script does not require an upward threshold crossing. If a feature stays above threshold, additional events may be accepted at the configured refractory spacing.

Event amplitude

Intensity mode: normalized = clamp((value - 40) / 40, 0, 1) Pitch mode: normalized = clamp((value - threshold) / threshold, 0, 1) Event amplitude = normalized × Amplitude_mapping Gaussian event = amplitude × exp(-((t - eventTime) / (effectiveDuration/6))²)

The same event-derived IR is written to every source channel. The IR duration equals the source duration. The detector is capped at 10,000 events.

Convolution and mix

A private zero-based copy is used for analysis and convolution; the selected source object is never modified.
The source is convolved with the event-driven IR using Praat Convolve: "sum", "zero".
The wet result is trimmed to the original duration.
The final Sound is built on a copy of the original source grid: wet × Wet + dry × (1-Wet).
Every non-silent result is then peak-normalized to 0.95.

Level behavior: the final Scale peak: 0.95 is target normalization, not an attenuation-only safety ceiling. Even a low-level result — including 0% Wet — is rescaled to a 0.95 peak if it is non-silent.

Visualization

The drawing contains:

Dry waveform and Result waveform, each drawn with Praat's own autoscaling.
Impulse panel with one stem/marker per accepted event and a dotted threshold reference.
Parameter line showing mode, threshold, spacing and event count.
Summary strip describing the self-derived IR relationship.

Marker heights are display-normalized feature values. They are not a direct plot of the Gaussian IR sample amplitudes.

Output behavior

Duration: exactly the original duration.
Time domain: final output is built on the original Sound grid, so the source time origin is retained.
Sample rate: preserved.
Channels: original channel count is preserved, including 3+ channel input.
Randomness: none; identical input and settings produce the same detector/IR.
No-event case: no output is created.