VIS (Vocal Insistunit Synthesizer)

Direct Synthesis Insitunit Generator for VocalTractLab

Generate vocal tract motor sequences without any audio input. Instead of analysing speech, this tool builds articulatory movement trajectories from scratch using a sequence of source–target pair gestures shaped by oscillator, chaotic attractor, and GENDY-style stochastic functions. The resulting base phrase — called the insitunit — is stored in memory and used as the template for any number of controlled variations, all concatenated into a single VTL tract-sequence file ready for synthesis.

---

Requirements

Package	Installation
numpy	pip install numpy
vocaltractlab-cython	pip install vocaltractlab-cython
write_reaper_markers	(optional) for REAPER marker embedding

Windows users: Python 3.10+ recommended. The scripts force UTF-8 stdout so Windows cp1252 terminals do not cause encoding errors.

---

Setup

Create the output folder from the script directory:

audio/output/

Place both files in the same directory:

directsynth_insitunit.py       ← backend / CLI
directsynth_insitunit_GUI.py   ← Tkinter GUI (run this)

---

Running

GUI (recommended)

python directsynth_insitunit_GUI.py

Command line

python directsynth_insitunit.py [options]

---

Core Concepts

Source–Target Pairs

The phrase is built from N source–target pairs. Each pair defines:

• A source position — a 30-dimensional vector of vocal tract / glottis parameter values
• A target position — another 30-dimensional parameter vector
• A duration in frames (1 frame = 110 audio samples ≈ 2.5 ms at 44 100 Hz)
• A movement generator — the function that shapes the path from source to target

In connected mode (default), the target of each pair becomes the source of the next, producing a continuous path through articulatory space.

Movement Generators

Each generator shapes how parameters travel from source to target over the pair's duration. The output is a [0, 1] curve that maps onto the source–target range.

Generator	Category	Description
linear	Smooth	Straight-line interpolation
sigmoid	Smooth	S-curve with configurable steepness
ease_in_out	Smooth	Smoothstep cubic — slow start and end
sine_arc	Smooth	Half-cosine arc — slow extremes, fast middle
elastic	Smooth	Overshoot-and-settle spring motion
logistic	Chaotic	Logistic map overlaid on a linear trend
lorenz	Chaotic	Lorenz attractor projected onto one axis
henon	Chaotic	Hénon map overlay
duffing	Chaotic	Duffing forced oscillator
ikeda	Chaotic	Ikeda optical bistability map
gendy	Stochastic	GENDY stochastic breakpoint synthesis (Xenakis)
random_walk	Stochastic	Biased random walk that arrives at the target

For chaotic generators, --chaos-depth (0–1) controls how much the attractor modulates around the smooth src→tgt trend. At 0.0 the chaos is silent; at 1.0 it dominates.

The Insitunit

The insitunit is the canonical base phrase — all pairs concatenated into a (30 × T) motor array. It is:

• Displayed as colour-coded parameter strips in the GUI Insitunit tab
• Saveable as a .npy file and reloadable in later sessions
• Used as the template for all subsequent variations

Variations

Variations are copies of the insitunit with controlled perturbation applied. Five perturbation methods are available:

Method	What it does
additive	Adds smooth interpolated noise scaled to a fraction of each parameter's range
scale_deviation	Scales each parameter's deviation from its centre value by ±depth
time_warp	Locally compresses and expands the time axis while preserving the sequence
stochastic	A stochastic scale envelope multiplies each parameter's deviation from centre
gendy_noise	GENDY-style stochastic breakpoint noise added to the trajectory

Scope controls which parameters are affected: all, tract (0–18), or glottis (19–29). A per-parameter JSON file can override the depth for individual parameters.

The output is: [insitunit] + [variation 1] + [variation 2] + … + [variation N] — all concatenated into one tract-sequence file.

---

GUI Overview

The GUI is organised into four tabs.

Phrase Design

• Number of source–target pairs and their frame duration range
• Position mode: random, center_biased (Gaussian around parameter centres), or extreme (biased toward range limits)
• Generator sequence: one row per pair; if the list is shorter than the number of pairs it cycles. Buttons for common presets (All GENDY, All Lorenz, Mix chaos)
• Generator knobs: chaos depth, GENDY breakpoints, amplitude step, duration step, GENDY distribution
• Estimated duration readout

Insitunit

• Generate now — builds the insitunit in-process and displays it immediately (no subprocess needed)
• Scrollable colour-coded thumbnail of all 30 parameter trajectories (tract = green, glottis = red)
• Save .npy / Load .npy for persistence across sessions

Variations

• Number of variations, global depth slider, scope and method selectors
• Method description panel updates to explain the selected method
• Per-parameter override table: enable/disable and set individual depth per parameter; bulk-set buttons for tract, glottis, or all

Output & Run

• Output directory selector
• Toggle audio synthesis (requires vocaltractlab-cython)
• Toggle REAPER marker embedding (requires write_reaper_markers)
• Use in-memory insitunit — pins the currently displayed insitunit so the run does not regenerate from scratch
• Log output panel; the subprocess runs in a background thread so the UI stays responsive

---

CLI Reference

Phrase Design

--num-pairs N           # Number of source-target pairs (default: 6)
--frames-min N          # Minimum frames per pair (default: 100)
--frames-max N          # Maximum frames per pair (default: 400)
--generators G [G ...]  # Generator(s); cycles if fewer than pairs (default: gendy)
--position-mode MODE    # random | center_biased | extreme (default: random)
--position-spread F     # Spread for center_biased mode, 0-1 (default: 0.3)
--no-connected          # Do not chain target -> next source
--seed N                # Random seed for reproducibility

Generator Parameters

--chaos-depth F          # Chaotic modulation depth, 0-1 (default: 0.25)
--gendy-breakpoints N    # GENDY breakpoint count (default: 8)
--gendy-amp-step F       # GENDY amplitude step size (default: 0.35)
--gendy-dur-step F       # GENDY duration step size (default: 0.25)
--gendy-distribution D   # cauchy | gaussian | uniform (default: cauchy)

Insitunit Persistence

--save-insitunit PATH    # Save generated insitunit to .npy
--load-insitunit PATH    # Load a saved .npy instead of generating

Variations

--num-variations N       # Number of variations (default: 0)
--variation-depth F      # Global depth, 0-1 (default: 0.15)
--variation-scope S      # all | tract | glottis (default: all)
--variation-method M     # additive | scale_deviation | time_warp |
                         # stochastic | gendy_noise (default: additive)
--per-param-config PATH  # JSON: {param_idx: depth} per-parameter overrides

Output

--output-dir PATH        # Output directory (default: audio/output)
--no-synthesize          # Write tract-sequence only; skip audio synthesis
--embed-markers          # Embed REAPER markers into the synthesized WAV

---

CLI Examples

# Minimal: 6 GENDY pairs, no variations, tract-sequence only
python directsynth_insitunit.py

# All Lorenz, 8 pairs, save the insitunit for reuse
python directsynth_insitunit.py \
    --num-pairs 8 \
    --generators lorenz \
    --frames-min 150 --frames-max 600 \
    --save-insitunit phrase_a.npy

# Load saved insitunit and generate 5 additive variations
python directsynth_insitunit.py \
    --load-insitunit phrase_a.npy \
    --num-variations 5 \
    --variation-depth 0.2 \
    --variation-method additive

# Mixed chaos generators, glottis-only gendy_noise variations
python directsynth_insitunit.py \
    --num-pairs 10 \
    --generators gendy lorenz henon duffing \
    --chaos-depth 0.4 \
    --num-variations 4 \
    --variation-depth 0.3 \
    --variation-scope glottis \
    --variation-method gendy_noise \
    --seed 42

# GENDY with Gaussian distribution and per-parameter override
python directsynth_insitunit.py \
    --generators gendy \
    --gendy-breakpoints 12 \
    --gendy-amp-step 0.5 \
    --gendy-distribution gaussian \
    --num-variations 3 \
    --per-param-config my_depths.json \
    --embed-markers

Per-parameter config JSON format:

{
  "0":  0.30,
  "1":  0.10,
  "19": 0.05,
  "20": 0.20
}

Keys are parameter indices (0–29). Parameters not listed are not varied.

---

Parameter Reference

Tract Parameters (indices 0–18):

HX  HY  JX  JA  LP  LD  VS  VO
TCX TCY TTX TTY TBX TBY TRX TRY
TS1 TS2 TS3

Glottis Parameters (indices 19–29):

f0  pressure  x_bottom  x_top  chink_area  lag
rel_amp  double_pulsing  pulse_skewness  flutter  aspiration_strength

---

Output

All files are written to audio/output/ (or --output-dir):

File	Description
*_tractseq.txt	VTL tract-sequence file — load in VocalTractLab 2.3
*.wav	Synthesized audio (requires vocaltractlab-cython)
*_markers.json	Insitunit / variation boundary timestamps
*_markers.wav	WAV with embedded REAPER cue markers (optional)

The tract-sequence file is compatible with VocalTractLab 2.3 — download at vocaltractlab.de. VocalTractLab runs on Windows only.

---

GENDY Notes

The GENDY generator is adapted from Iannis Xenakis's GENDY (GENerated DYnamics) stochastic sound synthesis algorithm (1992). In the original, a waveform is built from N breakpoints whose amplitude and duration are independently updated each cycle by random walks drawn from a probability distribution (Xenakis used Cauchy). Here the same mechanism is applied to parameter trajectories:

• Amplitude corresponds to the parameter's position within its source–target range
• Duration corresponds to the relative time width of each breakpoint segment
• The cauchy distribution (default) produces occasional large jumps separated by small steps — characteristic of Xenakis's original intent
• gaussian gives smoother, more bounded evolution
• uniform gives equal probability across the step range

The --gendy-breakpoints count controls the number of control points in the trajectory. Fewer breakpoints produce broader, slower shapes; more breakpoints produce finer texture.