Prosodic Reiterant Speech — KlattGrid Prosody Synth

Reiterant / "gibberish speech" synthesis using a continuous KlattGrid. Extracts pitch, intensity, and syllable timing from a source recording, then resynthesises it with a chosen consonant-vowel pattern (ma, la, na, ba, de) — preserving the source's prosodic contour while removing lexical content.

Author: Shai Cohen Affiliation: Department of Music, Bar-Ilan University, Israel Version: 2.6.1 (2026) License: MIT License Repo: https://github.com/ShaiCohen-ops/Praat-plugin_AudioTools

Contents:

What this does Quick start 5 Presets Syllable Patterns Controls Applications

What this does

This script implements reiterant / "gibberish speech" synthesis using a continuous KlattGrid. It extracts pitch, intensity, and syllable timing from a source recording, then resynthesises it with a chosen consonant-vowel pattern (ma, la, na, ba, de) — preserving the source's prosodic contour while removing lexical content. The result is a "gibberish" version of the speech that retains the original prosody (F0 contour, loudness envelope, rhythm) but replaces the phonetic content with a neutral, reiterant pattern.

What is reiterant speech? Reiterant speech is a technique used in prosody research where the original lexical content is replaced with a neutral syllable (e.g., /ma/ or /la/) while preserving the prosodic contour. This allows researchers to study prosody independently of segmental content. The script uses a continuous KlattGrid — a single synthesizer instance spanning the entire utterance — so the glottal source and formant filters run uninterrupted, producing natural-sounding speech without the "chunking" artifacts of per-syllable concatenation.

Key Features:

5 Presets — Default, Robotic, Natural Speech, Expressive, Minimal
7 Syllable Patterns — ma, la, na, ba, de_schwa, alternating, random
Speaker profiles — male, female, auto/broad
Naturalness continuum — 0 (robotic) to 1 (natural) with jitter, shimmer, OQ, formant drift
Pitch_strength — intonation gradient: 1 = natural contour, 0 = monotone at median
Consonant classes — nasal (/m/, /n/), lateral (/l/), plosive (/b/, /d/), with place cues
Research reproducibility — fixed seed for stochastic variation; 0 = unpredictable
Quality Control (QC) — F0 correlation, semitone RMSE, intensity shape RMSE, speech/pause agreement
Visualisation — source waveform with syllable boundaries, F0 overlay, intensity contour, output spectrogram/waveform, QC summary

v2.6 key improvements: Source Pitch dropouts no longer mute the synthetic signal frame-by-frame — V/UV gating is now opt-in and applies only to sustained unvoiced runs. Plosive closures are precomputed before the intensity envelope is built. Nasal pole+zero are explicitly controlled. QC separates source-vs-output F0 preservation from target-vs-output synthesis accuracy. The t=0 open-phase seed no longer collides with a real intensity-derived OQ point when speech begins at time zero.

Quick start

In Praat, select exactly one Sound object (speech recording).
Run script… → Prosodic_Reiterant_Speech.praat.
Choose Output_mode (reiterant_synth or lowpass_comparison).
Select a preset (Default, Robotic, Natural Speech, Expressive, Minimal, or Custom).
Choose Syllable_pattern (ma, la, na, ba, de_schwa, alternating, random).
Set Speaker_profile (male, female, auto/broad).
Adjust Naturalness (0 = robotic, 1 = natural) and Pitch_strength (1 = natural contour, 0 = monotone).
Click OK — the script analyses the source, builds a KlattGrid, synthesises, and outputs a new Sound.

Quick tip: Start with Natural Speech preset for a realistic reiterant synthesis. Robotic removes all stochastic variation (jitter, shimmer, formant drift) for a monotone, synthetic voice. Expressive maximises naturalness with high jitter and shimmer. Minimal is a fast, low-quality option for quick testing. Use Pitch_strength = 0 for a strictly monotone output at the median F0 (Parsons-style monotonisation).

Important: This script is pure Praat — no Python required. It uses Praat's KlattGrid synthesis engine (the same engine used for Klatt synthesizer speech synthesis). The source is analysed for pitch, intensity, and syllable timing; the output is a new Sound containing the reiterant speech. The script is experimental — perceptual intelligibility testing is recommended before use in research. The lowpass_comparison mode reuses the original waveform (filtered) and is intended as a baseline/comparison only.

5 Presets

Preset	Naturalness	Jitter	OQ	Silence (dB)	Min Pause (s)	Min Syll (s)	Max Syll (s)	Character
Default	1.0	0.4	0.70	25	0.12	0.08	0.45	Balanced natural synthesis
Robotic	0.0	0.0	0.50	25	0.12	0.08	0.45	No variation — flat, synthetic
Natural Speech	1.0	0.4	0.75	20	0.10	0.07	0.40	More natural, tighter timing
Expressive	1.0	0.8	0.85	18	0.08	0.06	0.50	High variation, expressive
Minimal	0.3	0.15	0.60	30	0.15	0.10	0.45	Fast, low-quality

Syllable Patterns — 7 Options

ma

Bilabial nasal + open vowel /a/. Nasal murmur with F2 locus ~1000 Hz.

la

Lateral approximant + /a/. F1 locus 250 Hz, F2 locus 1100 Hz, raised F3.

na

Alveolar nasal + /a/. Nasal murmur with F2 locus ~1750 Hz.

ba

Voiced plosive + /a/. Closure gap (35 ms) + VOT burst, F2 locus 800 Hz.

de_schwa

Alveolar plosive + schwa. F1 500 Hz, F2 1500 Hz, F2 locus 1800 Hz.

alternating

Cycles through ma, la, na, ba in order.

random

Randomly selects from ma, la, na, ba, de per syllable.

Consonant-class cues: Nasal consonants (/m/, /n/) use a nasal pole at ~270 Hz with a place-dependent antiformant (900 Hz for /m/, 1650 Hz for /n/). Plosives (/b/, /d/) use a brief closure gap (35 ms) followed by a VOT ramp and a formant locus (800 Hz for /b/, 1800 Hz for /d/). The lateral /l/ uses a smooth F1/F2 transition with a raised F3. These cues produce broad consonant-class distinctions rather than full phonetic naturalness.

Controls — Human/Robotic Continuum

Naturalness (0–1)

Master control for voice quality. 0 = robotic (all stochastic variation suppressed), 1 = natural (full jitter, shimmer, OQ variation, formant drift).

Jitter_amount (0–1)

Syllabic F0 microvariation depth. At Naturalness=1, 0.4 = ±1.6%, 0.8 = ±3.2%. Controls per-syllable pitch perturbation, not cycle-to-cycle jitter.

OQ_flat (0.3–0.9)

Baseline open quotient (glottal open phase). Higher = breathier voice. OQ swings with intensity (pressed voice at high intensity).

Pitch_strength (0–2)

Intonation gradient. 1 = natural contour from source, 0 = strictly monotone at median F0 (Parsons-style monotonisation). 0.5 = half the natural contour range.

Research_seed

Fixed seed for reproducibility. 0 = unpredictable run; non-zero = deterministic (same random variation each run).

Gate_long_unvoiced_runs

Opt-in V/UV gating. If enabled, only contiguous source-unvoiced runs longer than Minimum_unvoiced_gate_s (default 50 ms) can mute the synthetic signal. Prevents short tracker dropouts from punching holes in the output.

Naturalness scaling

Jitter: max ±4% × jitter_amount × naturalness

Shimmer: max ±7% × naturalness (amplitude perturbation)

OQ swing: 0.15 × naturalness (intensity-driven open quotient variation)

Formant drift: F1: ±3% × naturalness, F2: ±2% × naturalness

Applications

Prosody research stimuli

Use case: Generate reiterant speech stimuli for prosody perception experiments. The output preserves the source's F0 contour, intensity envelope, and rhythm while removing lexical content.

Settings: Natural Speech preset, Pitch_strength=1.0. The output is a "gibberish" version of the source that can be used to test prosodic perception without segmental cues.

Monotonisation control (Parsons 2025)

Use case: Create a monotone version of a speech utterance at the speaker's median F0.

Settings: Pitch_strength=0, Robotic preset. The output is a flat, monotone synthesis — useful as a baseline for prosody manipulation studies.

Low-pass comparison baseline

Use case: Generate an F0-individualised low-pass filtered version of the source (Parsons et al., Interspeech 2025).

Settings: Output_mode=lowpass_comparison. The cutoff is 420.2 × (1 − e^{−0.0124 × meanF0}) with transition bandwidth = meanF0/4. Reuses the original waveform (filtered).

Expressive / robotic voice continuum

Use case: Explore the human-to-robotic voice quality continuum for sound design or synthesis research.

Settings: Adjust Naturalness from 0 to 1, with Expressive preset for maximum variation. The output ranges from flat, robotic speech to expressive, natural-sounding gibberish.

Workflow: Voice recording → Reiterant speech

Source: Spoken sentence.
Settings: Natural Speech preset, pattern=alternating (ma, la, na, ba).
Result: The source's prosody is preserved, but the lexical content is replaced with alternating CV syllables — a "gibberish" version of the original utterance.

Workflow: Monotonisation control

Source: Any speech recording.
Settings: Robotic preset, Pitch_strength=0.
Result: The output is a flat, monotone synthesis at the speaker's median F0 — no intonation contour.

Workflow: Low-pass comparison

Source: Speech recording.
Settings: Output_mode=lowpass_comparison.
Result: A low-pass filtered version of the original with an F0-individualised cutoff. Intended as a comparison/baseline condition for prosody obfuscation studies.

Troubleshooting:
• No speech detected: Lower Silence_threshold_dB_below_peak (e.g., 25 → 18). If the recording is very quiet, the script may find no speech regions.
• Output has clicks or discontinuities: The continuous KlattGrid should avoid the "chunking" artifacts of per-syllable concatenation. If clicks occur, check that the source duration is long enough (>0.5 s) and that the pitch range is appropriate.
• Pitch tracking fails: Adjust Pitch_floor and Pitch_ceiling to match the speaker. For female voices, use floor=100, ceiling=500. For male voices, floor=75, ceiling=300.
• Naturalness not audible: Set Naturalness=1 and use the Expressive preset. The stochastic variation (jitter, shimmer, OQ, formant drift) may be subtle; increase Jitter_amount to 0.8 for more noticeable variation.
• Plosive closure not audible: The closure gap is 35 ms by default. For longer closures, edit the plosClosed constant in the script (currently 0.035 s).
• Output is too quiet: Increase Output_peak (0.85 → 0.95). The final Scale peak sets the global output level.

Quality Control (QC)

When Draw_QC is enabled, the script generates a multi-panel scientific figure:

Panel A: Source waveform with syllable boundaries (dotted) and nuclei (ticks).
Panel B: F0 (Hz) — source vs output overlay, with median reference.
Panel C: Intensity contour — source vs output shape (global level removed).
Panel D: Output spectrogram (with F0 overlaid) or output waveform.
Summary strip: F0 preservation (source-output correlation), synthesis accuracy (target-output correlation + RMSE), intensity shape RMSE, speech/pause agreement, seed.

The QC metrics are also printed in the Info window: source/output F0 correlation, target/output F0 correlation and semitone RMSE, intensity contour correlation and shape RMSE, speech/pause agreement.