Markov Soundscape Weaver — User Guide

What this does

This script analyzes one Sound as a sequence of overlapping acoustic grains, groups those grains into texture states with k-means, and models state-to-state transitions with either a first-order or second-order Markov chain. A new state sequence is then generated from the learned transition model, and each state selects source grains for resynthesis.

Key Features:

• Overlapping analysis grains — analysis hop is determined by Analysis_overlap
• 4D acoustic feature space — spectral centroid, spectral bandwidth, pitch, HNR
• Per-feature z-score normalization
• k-means texture states with empty-cluster reseeding and final pruning
• First- or second-order Markov chains
• Randomness control — explicit probability of bypassing the learned transition and jumping by state occupancy
• State-conditioned grain emission
• Source-position jitter and output-onset density variation
• Varispeed scatter — pitch, speed, and grain duration change together
• Stereo decorrelation continuum from identical event lists to fully independent walks
• Normalized arbitrary-overlap OLA with accumulated analytic envelope
• Exact requested output duration
• Random_seed for reproducibility
• Suite-standard visualization

Technical Implementation: (1) Convert the source to mono. (2) Extract overlapping 4D analysis grains. (3) Z-score each feature independently. (4) Run k-means with empty-state reseeding. (5) Build either a first-order transition matrix or a second-order transition tensor with first-order backoff. (6) Generate one event list containing state, source grain, source position, varispeed ratio, and output onset. (7) Derive the right-channel event list according to Stereo_decorrelation. (8) Render each channel into audio and envelope buffers. (9) Divide by the accumulated envelope, trim to the exact duration, apply boundary fades, combine stereo if required, and peak-scale to 0.99.

Quick start

1. Select exactly one Sound object in Praat.
2. Run Markov_Soundscape_Weaver.praat.
3. Choose a preset or Manual.
4. For Manual, set Grain_size_ms, Analysis_overlap, Number_of_states, Markov_order, Randomness, Output_duration_sec, Synthesis_overlap, Varispeed_scatter_semitones, Position_jitter, Density_variation, Stereo_decorrelation, and Random_seed.
5. Enable visualization if you want to inspect the generated state path, transition matrix, state populations, and waveform comparison.
6. Click OK. Output is named <source>_MarkovWeave_<preset>.

Markov Theory

First-Order Chain

Observed state-to-state transitions in the source are counted and each row is normalized. Empty target states are excluded. A row with no usable observed transitions falls back to a distribution over active states.

Second-Order Chain

The second-order model uses a K×K context pair and a successor state. Generation is seeded from a state pair that actually occurred in the source sequence. If a context pair was never observed, the script backs off to the learned first-order distribution conditioned on the current state. If that row is also unavailable, active-state occupancy is used.

Randomness

Therefore Randomness = 0 follows the learned grammar exclusively, while higher values progressively inject occupancy-weighted state jumps.

Initial State

First-order generation starts from an active state sampled by state occupancy. Second-order generation starts from an observed adjacent state pair rather than two unrelated random states.

Grammar Time Scale

If the two overlaps differ, the learned sequence is played at a different event rate. A ratio below 1 compresses the grammar in time; above 1 expands it. This is an intentional compositional parameter interaction, not an error.

Analysis Phase

Overlapping Grain Grid

Analysis_overlap is clamped to 0–0.95. The source must be at least four grain lengths long and must provide at least twice as many analysis grains as requested states.

4D Feature Space

Z-Score Normalization

K-Means Texture States

After clustering, any state that somehow remains empty is pruned from the Markov chain: it cannot be an initial state, successor state, or grain-emitting state.

Synthesis Phase

State-Conditioned Grain Selection

Position_jitter therefore changes where the source is read. It does not change the output onset.

Varispeed Scatter

This is varispeed: pitch, playback speed, and grain duration change together.

Density Variation

Density_variation changes the output event timing. Position_jitter changes the source read position. They are different controls.

Stereo Decorrelation

Normalized Overlap-Add Rendering

The renderer supports arbitrary overlap, density jitter, and variable grain durations after varispeed. The old block-based concatenate system no longer exists.

Final Output

Parameters & Presets

Manual Parameters

Script-Level Edge Taper

Built-in Presets

Visualization

v0.8 Diagnostic Page

Applications

Texture-State Resequencing

Generate new arrangements of a source's acoustic texture states while retaining transition tendencies learned from that source.

Soundscape Variation

Environmental, instrumental, vocal, or mixed recordings can produce useful generative textures when their local spectral/pitch/HNR patterns form recurring state transitions.

First- vs Second-Order Grammar Experiments

Compare the same source under first-order and second-order transition models to hear how conditioning on the previous two states changes local continuity.

Temporal Rescaling of Learned Grammar

Changing Analysis_overlap and Synthesis_overlap independently changes the event-rate relationship between learned and rendered sequences without altering the source feature model itself.

Stereo Texture Design

Stereo_decorrelation provides a controlled continuum from duplicated mono-in-stereo to independent probabilistic walks.

Troubleshooting