LPC Voice Morphing — LPC Vocoder Pro — User Guide

An LPC excitation-replacement vocoder that filters pitch-driven phonation or Gaussian noise through the time-varying LPC envelope of the selected source.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel Version: 0.5 (2026) License: MIT License Repo: https://github.com/ShaiCohen-ops/Praat-plugin_AudioTools

Contents:

What this does LPC vocoder / source–filter model Quick start Presets Controls Processing pipeline Channels & level Visualization Historical / compositional context Further reading

What this does

LPC Voice Morphing / LPC Vocoder Pro is an excitation-replacement vocoder. It analyzes one channel of the source with LPC, creates either a pitch-driven phonation source or Gaussian noise, and filters that new excitation through the source's time-varying LPC envelope.

Despite the historical file name, this is not a morph between two voices or between two sets of LPC coefficients. The “morph” is the replacement of the excitation while retaining the analyzed spectral envelope of one source.

LPC vocoder and source–filter model

LPC treats the short-time spectrum as the response of an all-pole filter driven by an excitation. In speech terms, the filter approximates the vocal-tract resonances; the excitation can be periodic, noisy, or synthetic. A vocoder becomes creative when those components are deliberately separated and recombined.

This script offers two excitation families:

Pitch-driven pulse: Praat extracts pitch, converts it to a PitchTier, optionally reshapes it, and generates a phonation waveform.
Gaussian noise: broadband random excitation produces whisper-like or noise-vocoder textures while the LPC filter supplies the source's changing resonant envelope.

Continuous pitch bridge: on the pulse path, the PitchTier deliberately bridges unvoiced gaps. This is a musical design choice: the synthetic excitation can remain continuous where stricter speech resynthesis would switch between voiced and unvoiced sources.

Quick start

Select exactly one Sound.
Run LPC_Voice_Morphing.praat.
Try Natural Resynthesis for the core LPC character, Robot for monotone excitation, or Whisper for noise excitation.
Use Custom when you need direct control of excitation type, pitch range, LPC method/order and target intensity.

Presets

Preset	Excitation	Window	Pitch behavior	Other override
Natural Resynthesis	Pitch-driven pulse	25 ms	Source PitchTier, continuous across gaps	LPC order Auto; Force monotone OFF
Robot Voice	Pitch-driven pulse	30 ms	Forced to 100 Hz	LPC order Auto
Whisper	Gaussian noise	15 ms	Pitch controls are not used for excitation	LPC order Auto; recursive brightness stage after synthesis
Deep Demon	Pitch-driven pulse	40 ms	PitchTier × 0.6; minimum pitch forced to 50 Hz	LPC order Auto; Force monotone OFF

Presets do not overwrite every form field. In particular, LPC method, pre-emphasis, target intensity, playback and visualization remain user choices; most pitch-analysis values also remain available except where a preset explicitly changes them.

Controls

Control	Meaning
Excitation source	Pitch-driven phonation or Gaussian noise. Used directly in Custom; presets may override it.
Time step	Pitch-analysis step used on the pulse path.
Minimum / Maximum pitch	Pitch-search range. Maximum must exceed minimum.
Force monotone	Flattens the PitchTier to the chosen monotone frequency on the pulse path.
Monotone frequency	Frequency used for Force monotone and also as the fallback when no pitch points are detected.
LPC method	Legacy autocorrelation character or Burg. Autocorrelation automatically falls back to Burg only if the render becomes numerically catastrophic.
LPC order	0 = Auto. Auto uses `round(sampleRate / 1000) + 4`.
Analysis window	LPC analysis window in seconds. LPC frame step is fixed at 5 ms.
Pre-emphasis Hz	Pre-emphasis setting passed to Praat LPC analysis.
Target intensity dB	Praat intensity target in dB SPL, not dBFS.

Processing pipeline

Choose analysis driver: mono input is copied; for multichannel input the highest-RMS source channel is selected instead of summing channels.
Create excitation: either pitch-driven phonation at the source sample rate or Gaussian noise over the source duration.
Analyze LPC: autocorrelation or Burg, using the selected analysis driver.
Vocode: filter the synthetic excitation through the time-varying LPC model.
Stability fallback: if legacy autocorrelation output becomes undefined or exceeds the catastrophic peak threshold, rerender automatically with Burg.
Level: if the analyzed source is not effectively silent, scale to the requested Praat intensity.
Whisper only: apply the retained recursive brightness formula.
Safety: attenuate to 0.99 only when the final peak exceeds 0.99.

Channels, duration and level

The vocoder output is mono, even when the source is stereo or multichannel.
For multichannel input, the highest-RMS channel alone drives pitch and LPC analysis; this avoids anti-phase cancellation from a stereo sum.
Pulse and noise excitation use the original duration and source sample rate.
Target intensity uses Praat's dB SPL convention.
The final 0.99 stage is a down-only safety ceiling, not target peak normalization.
If the analysis driver is effectively silent, the final output is forced to silence rather than allowing the fallback excitation to create a tone from nothing.

Visualization

Original and vocoded waveforms are drawn on one shared amplitude scale. The guide also shows original and vocoded spectrograms up to 4 kHz and reports the realized LPC order/method, analysis channel, window, excitation type and target level. If autocorrelation fell back to Burg, the summary reports the method actually used.

Historical, technological and compositional context

The processor sits at the intersection of two related traditions. The first is the vocoder: Homer Dudley's Bell Labs work in the 1930s demonstrated that speech could be analyzed into slowly varying spectral information and reconstructed from an artificial buzz or noise source. The second is linear prediction, developed for efficient speech representation and formalized in influential analysis–synthesis work by Atal and Hanauer and in Makhoul's 1975 tutorial review.

Compositionally, vocoding turns the source–filter model into a way of separating identity from excitation. A pitch contour can be flattened, lowered or replaced by noise while the resonant envelope still carries recognizable traces of the original articulation. The Robot, Whisper and Deep Demon presets exploit exactly that separation. The processor is therefore closer to a creative LPC vocoder than to transparent voice conversion.