LPC Voice Morphing — User Guide

What this does

This script implements an LPC (Linear Predictive Coding) vocoder — a classic speech analysis‑synthesis technique that models the human vocal system. LPC decomposes speech into two components: (1) Excitation source (glottal pulses for voiced sounds, noise for unvoiced/whispered sounds), representing pitch and breath; and (2) Spectral envelope (vocal tract filter), representing formants and timbre. The script extracts these components from an input sound, allows modification of the excitation (e.g., making pitch monotone for robot voice, replacing with noise for whisper), then re‑filters the modified excitation through the original spectral envelope. The result preserves linguistic content (phonemes, words) while transforming vocal quality (pitch, breathiness, timbre).

Key Features:

• 4 Themed Presets — Natural Resynthesis, Robot Voice (Monotone), Whisper (True Noise), Deep Demon
• Dual Excitation Sources — Pulse train (voiced) or Gaussian noise (unvoiced/whisper)
• Auto‑calibrated LPC Order — Calculates optimal filter complexity based on sample rate
• Pitch Manipulation — Force monotone pitch, shift pitch (Deep Demon), or preserve natural pitch
• True Spectral Envelope Extraction — LPC analysis captures formant structure accurately
• Post‑processing Optimization — Intensity scaling, whisper brightness correction
• Non‑destructive Processing — Original sound preserved, output named with "_vocoded" suffix

Technical Implementation: The script follows the classic LPC vocoder pipeline: (1) Pitch analysis (Praat's Pitch object) or noise source creation. (2) Excitation generation — either pulse train via Praat's "To Sound (phonation)" with LF model parameters (7‑argument version) or Gaussian noise. (3) LPC analysis — Praat's "To LPC (autocorrelation)" extracts filter coefficients from original sound. (4) Filtering — Modified excitation is filtered through LPC filter (all‑pole synthesis). (5) Post‑processing — Intensity normalization, high‑frequency enhancement for whispers. The script handles the complex object conversions (Sound → Pitch → PitchTier → Pitch → PointProcess → Sound) required for proper excitation generation.

Quick start

1. In Praat, select exactly one Sound object containing speech.
2. Run script… → lpc_vocoder_pro.praat.
3. Choose a Preset (Natural, Robot, Whisper, Deep Demon) or Custom.
4. Adjust source parameters if needed (time_step, pitch range, monotone options).
5. Adjust LPC filter parameters if needed (order, window, pre‑emphasis).
6. Set target_intensity_db (default 70 dB SPL) and play_after_processing.
7. Click OK — script analyzes, generates excitation, filters, normalizes.
8. Output appears as: originalName_vocoded.
9. Watch Info window for progress (excitation generation, LPC analysis, vocoding).

LPC Theory & Source‑Filter Model

The Source‑Filter Model of Speech

Linear Predictive Coding Mathematics

LPC vs. Other Speech Analysis Methods

Complete LPC Vocoder Pipeline

INPUT: Original speech sound

STEP 1: PITCH ANALYSIS (for voiced excitation)
  • To Pitch (Praat): time_step, min_pitch, max_pitch
  • Optionally modify: force_monotone, pitch_shift
  • Convert to excitation: Pitch → PitchTier → Pitch → PointProcess → Sound

STEP 2: NOISE GENERATION (for whisper)
  • Create Gaussian noise: randomGauss(0, 0.2)
  • Same duration as original

STEP 3: LPC ANALYSIS (spectral envelope)
  • To LPC (autocorrelation): order, window, pre‑emphasis
  • Computes coefficients {aₖ} and gain G per frame
  • Captures formant structure

STEP 4: FILTERING (synthesis)
  • Filter source through LPC filter
  • Praat's "Filter" command with LPC object
  • Result: source shaped by original's spectral envelope

STEP 5: POST‑PROCESSING
  • Scale intensity to target_db
  • Whisper: high‑frequency emphasis
  • Output naming

OUTPUT: originalName_vocoded
    

Preset Characters

Preset 1: Custom

Preset 2: Natural Resynthesis

Preset 3: Robot Voice (Monotone)

Preset 4: Whisper (True Noise)

Preset 5: Deep Demon

Preset Comparison Table

Excitation Source Generation

Voiced Source: LF Model Pulse Train

LPC Filter Analysis

LPC Order Determination

Praat's LPC Analysis Parameters

Spectral Envelope Extraction

Post‑processing: Whisper Brightness Correction

Parameters & Controls

Form Parameters

Parameter Guidelines

Internal Variables & Objects

Sonic Applications

Voice Design for Media

Music Production

Vocal processing:

• Robot backing vocals: Apply Robot preset to harmony parts
• Whispered doubles: Create whisper version to layer under main vocal
• Pitch‑corrected effects: Use Natural preset but with modified pitch contour (custom PitchTier formula)
• Vocal synthesis from instruments: Process instrumental sounds through LPC trained on speech (experimental)

Speech Technology & Education

Speech analysis demonstrations: Show source‑filter separation by listening to excitation and filtered result separately.

Accent/voice modification: Process speech while preserving linguistic content but altering voice characteristics.

Privacy masking: Whisper preset obscures speaker identity while preserving words (for interviews, sensitive content).

Speech compression understanding: Demonstrate how LPC achieves high compression (few coefficients represent speech).

Practical Workflow Examples

Advanced Techniques & Customization

Custom Pitch Manipulation

Modify PitchTier formula for custom effects:

Mixed Excitation (Voiced/Unvoiced)

Create more natural synthesis with voicing decisions:

LPC Coefficient Manipulation

Modify formant structure directly: After LPC analysis, you could modify coefficients before filtering:

Cross‑Synthesis (Voice Transfer)

Use one sound's excitation with another's LPC filter:

Troubleshooting