LPC Excitation Lab — User Guide

Offline LPC analysis–resynthesis for driving a source sound’s spectral envelope with synthetic or cross-synthesized excitation.

Author: Shai Cohen Version: 2.2.1 (2026) Application: Praat scripting language License: MIT License

Contents:

What this does Signal flow Quick start Cross-synthesis modes Excitation methods Presets Parameters Input & output behavior Visualization Notes & limitations Further reading

What this does

LPC Excitation Lab separates the roles of excitation, spectral envelope, and articulation. It analyzes a source with linear predictive coding (LPC), builds or extracts an excitation signal, inverse-filters that excitation with its own LPC model to obtain a residual, then filters the residual through the LPC model chosen as the destination spectral envelope. Finally, it applies an intensity contour as the articulation envelope.

The result is not a conventional pitch shifter or formant editor. The spectral coloration is imposed by an LPC all-pole model, while the temporal fine structure comes from the selected excitation. In cross-synthesis modes, different Sounds can provide these roles.

Core idea: excitation determines the fine-grained source character; the destination LPC determines the time-varying resonant envelope; the selected IntensityTier determines the broad amplitude articulation.

Signal flow

Excitation
   ↓
LPC analysis of excitation
   ↓
Inverse LPC filter → residual
   ↓
Destination LPC filter
   ↓
IntensityTier multiplication
   ↓
Peak scale to 0.99
   ↓
Mono output

The destination LPC is derived with Praat’s To LPC (burg) using a fixed 25 ms analysis window, 5 ms frame step, 50 Hz pre-emphasis frequency, and the user-selected LPC order.

Quick start

Select one Sound for Off (single Sound), or exactly two Sounds for either cross-synthesis mode.
Run LPC_Excitation_Lab.praat.
Choose a preset or Custom.
Choose the Cross_synth_mode. Preset Cross-Synth automatically selects mode 2.
For modes using synthetic excitation, choose an excitation method and adjust only the parameters relevant to that method.
Set Lpc_order. Higher orders allow a more detailed all-pole envelope, but can also make the model more sensitive to the material.
Optionally enable the visualization and playback.

Selection order matters in cross-synthesis. Sound1 and Sound2 have different roles. The tables below describe exactly which Sound supplies the residual, filter, and intensity contour.

Cross-synthesis modes

Mode	Excitation / residual	LPC filter	Intensity contour	Duration
1. Off (single Sound)	Synthetic excitation selected in the form	Sound1	Sound1	Sound1 duration; Time Stretch can extend it
2. Sound2 residual through Sound1 filter	Sound2 is LPC-inverse-filtered; synthetic excitation controls are ignored	Sound1	Sound2	`min(duration1, duration2)`
3. Synth excitation through Sound2 filter	Synthetic excitation selected in the form	Sound2	Sound1	`min(duration1, duration2)`

Mode 2 is the direct cross-synthesis configuration: Sound2 contributes residual/fine temporal structure and articulation, while Sound1 contributes the LPC spectral envelope. Mode 3 keeps a synthetic excitation but substitutes Sound2’s LPC envelope; Sound1 supplies only the articulation contour.

Excitation methods

1. Pitch Sweep

A Praat PitchTier drives a pulse-train source whose base frequency moves linearly in Hz from Start_freq_hz to End_freq_hz. The tier is sampled every 5 ms so optional vibrato and random-walk pitch jitter can be added.

Vibrato_rate_hz and Vibrato_depth_cents add sinusoidal vibrato.
Pitch_jitter_cents adds a bounded, mean-reverting random walk in cents.

2. Pulse Train

Builds a PointProcess with nominal period Period_s, then converts it to a pulse-train Sound.

Pulse_jitter_pct randomizes each period around the nominal value.
Pulse_shimmer_pct applies a deterministic two-sine amplitude modulation after pulse generation; it is a shimmer-like amplitude wobble, not a measured acoustic shimmer parameter.

3. Time Stretch

Uses Gaussian noise as the synthetic excitation. In single-source mode only, the destination LPC time axis and the Sound1 intensity contour are scaled by Time_stretch_factor, so the output duration becomes approximately the source duration multiplied by that factor.

Noise_am_rate_hz and Noise_am_depth optionally modulate the noise amplitude with a sinusoid. In cross-synthesis mode 3, Time_stretch_factor is ignored; the mode uses the normal cross-synthesis duration.

4. Chirp Burst

Generates one or more exponential chirps between Chirp_start_hz and Chirp_end_hz. Additional layers multiply both endpoints by values parsed from Layer_ratios. Up to eight layers are used; frequencies are clamped below 95% of Nyquist. If the requested layer count exceeds the number of supplied ratios, the parser falls back to the last available ratio.

5. Granular Noise

Builds a noise excitation from Gaussian-windowed grains. The nominal grain size is 40 ms; Grain_size_modulation_pct adds per-grain random size variation. Density changes linearly from Grain_density_start to Grain_density_end, with grain spacing calculated as 40 ms / density. Thus higher density means more frequent grains.

Presets

Presets override the parameters listed below. Other form fields retain their current values, so a preset does not necessarily reset unrelated modulation controls.

Preset	Overrides
Custom	No preset override.
Voiced Sweep	Pitch Sweep; 80 → 400 Hz; LPC order 46.
Robotic Pulse	Pulse Train; period 8 ms; LPC order 46.
Stretched Whisper	Time Stretch; factor 5.0; LPC order 50.
Chirp Texture	Chirp Burst; 60 → 8000 Hz; LPC order 46.
Grain Cloud	Granular Noise; density 0.4 → 0.4; LPC order 40.
Vibrato Voice	Pitch Sweep; 120 → 220 Hz; 5.5 Hz vibrato; ±35 cents vibrato depth; ±8 cents pitch jitter; LPC order 46.
Jittered Pulse	Pulse Train; 7.3 ms period; ±8% period jitter; ±25% shimmer-like amplitude wobble; LPC order 46.
Density Cloud	Granular Noise; density 0.2 → 0.9; ±30% grain-size modulation; LPC order 42.
Chirp Stack	Chirp Burst; 80 → 3000 Hz; 4 layers; ratios 1.0, 1.5, 2.7, 4.2; LPC order 50.
Cross-Synth	Cross-synthesis mode 2; LPC order 48. The displayed excitation method is irrelevant because mode 2 uses Sound2’s residual.

Parameters

Parameter	Default	Meaning
Cross_synth_mode	Off	Selects the assignment of excitation/residual, LPC filter, and articulation source.
Excitation_method	Pitch Sweep	Synthetic source used in modes 1 and 3. Ignored in mode 2.
Start_freq_hz / End_freq_hz	80 / 300 Hz	Pitch Sweep endpoints.
Vibrato_rate_hz / depth	0 / 0	Pitch Sweep vibrato; depth is in cents.
Pitch_jitter_cents	0	Random-walk pitch deviation for Pitch Sweep.
Period_s	0.01 s	Nominal Pulse Train period.
Pulse_jitter_pct	0	Random period variation.
Pulse_shimmer_pct	0	Depth of the post-generation amplitude wobble.
Time_stretch_factor	3.0	Single-source Time Stretch only.
Noise_am_rate_hz / depth	0 / 0	Sinusoidal amplitude modulation of Time Stretch noise.
Chirp_start_hz / end_hz	100 / 4000 Hz	Base exponential-chirp endpoints.
Chirp_layers	1	Number of stacked chirps, clamped to 1–8 internally.
Layer_ratios	1.0, 1.5, 2.0, 2.7	Multipliers applied to chirp start/end frequencies.
Grain_density_start / end	0.6 / 0.6	Start/end grain-rate control, internally constrained to 0.01–1.
Grain_size_modulation_pct	0	Per-grain random size variation.
Lpc_order	46	Prediction order for both the destination LPC and the excitation LPC.
Draw_visualization	On	Draw the Praat AudioTools diagnostic page.
Play_result	On	Play the output after processing.

Fixed analysis settings: LPC window = 25 ms; frame step = 5 ms; pre-emphasis frequency = 50 Hz. These are not exposed in the form.

Input & output behavior

Selection: exactly one Sound in mode 1; exactly two Sounds in modes 2 and 3.
Minimum duration: at least 64 ms of effective material.
Sample rate: preserved from Sound1. In two-Sound modes, both Sounds must have exactly the same sample rate.
Channels: LPC processing and output are always mono. Multichannel inputs are averaged to mono before analysis/resynthesis.
Duration: mode 1 normally preserves Sound1 duration. Single-source Time Stretch multiplies the duration by Time_stretch_factor. Modes 2 and 3 use the shorter of Sound1 and Sound2.
Start time: internal processing uses a 0-based working domain; the final output is shifted back to Sound1’s original xmin.
Output level: after intensity multiplication, the final Sound is always peak-scaled to 0.99. There is no Natural-level or bypass output mode.
Output name: Sound1_Method_Preset. In mode 2 the method label is Sound2Residual.

Intensity multiplication: the articulation contour is measured before synthesis. Mode 2 uses Sound2’s intensity; modes 1 and 3 use Sound1’s. In single-source Time Stretch, the IntensityTier is stretched with the LPC time axis.

Visualization

When Draw_visualization is enabled, the 8-inch Praat AudioTools page contains:

Excitation waveform — the synthetic excitation, or Sound2 in residual cross-synthesis mode.
LPC residual waveform — the excitation after inverse filtering by its own LPC model.
Output waveform — the re-filtered signal after intensity articulation.
Spectrum overlay — matched 80 ms Hanning-windowed excerpts from the destination filter source and output, displayed from 0–8 kHz.
Summary strip — preset, method, cross-synthesis mode, LPC order, analysis settings, source durations, output duration, peak, and sample rate.

The spectrum overlay is a local comparison, not a plot of the LPC coefficients themselves. For equal-duration material, both spectra use the same midpoint interval. For Time Stretch, the same 80 ms window length is taken at the relative midpoint of each signal.

Notes & limitations

LPC is an all-pole model. The result is therefore shaped by what that model can represent; it is not an exact spectral transplant.
LPC order is not a formant count. This tool uses LPC as a creative time-varying spectral envelope, not as a formant-measurement command.
Randomness: Time Stretch noise and Granular Noise are random on every run. Pitch jitter and pulse jitter also introduce randomness when nonzero. There is no random-seed control, so such settings are not reproducible sample-for-sample across runs.
Pulse shimmer is synthetic modulation. It is not calculated from the source and should not be interpreted as a measured shimmer statistic.
Time Stretch factor is ignored in mode 3. Mode 2 does not use synthetic excitation controls at all.
Very high LPC orders or unsuitable material can produce strongly resonant or unstable-sounding resynthesis even when the script completes normally.
Peak normalization is unconditional. Level comparisons with the source should therefore be made after accounting for the final 0.99 scaling.