Latent Pbind — Algorithmic Latent-Space Composition

What this does

Latent Pbind trains a compact beta-VAE on spectral patches extracted from one selected Sound, normalises the learned latent space into musically usable coordinates, and lets a Pbind-style event language generate an offline trajectory through that space. Each latent coordinate can be driven by a separate deterministic, stochastic, cyclic, Brownian, chaotic, or nested pattern.

Key Features:

• On-the-fly latent learning — a small NumPy beta-VAE is trained directly from the selected Sound.
• Pbind-style composition — independent pattern streams control z1...z8, cluster identity, amplitude, temperature, attraction, source-read rate, and retriggering.
• Independent clocks — dur_z1, dur_z2, etc. let latent dimensions behave as separate temporal voices.
• Per-key interpolation — Step, Linear, Cubic, or Smooth globally or independently per stream.
• Three rendering modes — spectral-envelope Decoder, Nearest event, and Barycentric resynthesis.
• Four boundary behaviors — Clamp, Reflect, Wrap, and Soft attract.
• Robust latent coordinates — PCA rotation and percentile scaling map the learned data to stable approximately -1...+1 units.
• Identity clusters — k-means++ centroids provide macro-level acoustic regions that can be combined with local latent offsets.
• Reproducible stochastic structure — every pattern key has its own seeded random stream.
• Explanatory Praat Picture visualisation — pattern events, latent trajectories, latent maps, rendering diagnostics, output waveform, and summary.
• CPU-only and offline — no pretrained model, cloud service, API, GPU, PyTorch, or TensorFlow.

Quick start

1. Select exactly one Sound object in Praat.
2. Run LatentPbind.praat.
3. For a first run, choose Preset = Polymetric coordinates or Brownian latent walk.
4. Keep Rendering mode = Decoder, Interpolation = Smooth, and Boundary mode = Reflect.
5. Leave Output duration = 0 to use the input duration.
6. Keep Excursion = 1.0 so the pattern works in the robustly normalised data range.
7. Enable Draw visualisation to inspect the event streams and their movement through the learned latent space.

Core concept

Latent Pbind separates the control source from the learned acoustic space. The control source is a compact Pbind-style expression; the target is a latent representation learned from the selected Sound.

The central compositional idea is that latent dimensions can behave like independent voices. One coordinate may drift slowly, another may repeat an asymmetric cycle, a third may make sparse jumps, and a fourth may decay geometrically. Their interaction defines one multidimensional path through the learned sound space.

Pbind language

The parser accepts a compact subset of SuperCollider-inspired pattern syntax. Both keyword style and backslash-key style are accepted.

Pbind(
    dur=Pseq([0.25,0.5,0.25,1],inf),
    z1=Pwhite(-1,1,inf),
    z2=Pwalk([-1,-0.5,0,0.5,1],1),
    z3=Pseq([0,0.5,1,-0.5],inf),
    interp=cubic
)

Equivalent key/value style is also accepted:

Pbind(\dur, Pseq([0.25,0.5],inf), \z1, Pbrown(-1,1,0.2,inf))

Supported patterns

Pbind keys

Latent model

1. Spectral patches

The input is converted to mono and analysed with an STFT. Each frame is reduced to 64 log-spaced power bands. Overlapping patches are then resampled to 8 time cells, producing a 512-value representation for each non-silent patch.

2. NumPy beta-VAE

A compact fully connected beta-VAE is trained from scratch:

The model uses tanh hidden layers, Adam optimisation, and KL warm-up. The default latent dimensionality is 4 and can be set from 2 to 8.

3. Rotation and robust normalisation

The posterior means are PCA-rotated so z1 is the direction of largest spread, followed by z2, etc. Each axis is sign-oriented so its correlation with spectral centroid is non-negative; this orientation is musically meaningful only when the reported correlation is clearly above zero.

The 2nd and 98th percentiles of each rotated coordinate define the normalised range approximately -1...+1. This prevents a single outlier from defining the user-facing scale.

4. Identity clusters

k-means++ clusters the normalised latent cloud. Cluster centres are ordered along z1, so cluster 0 is the centroid with the lowest z1. A cluster pattern can therefore provide a macro-level identity trajectory while z1...zd provide local offsets around it.

7 Presets + Custom

Controls

Advanced settings

Navigation & boundaries

The raw trajectory is formed as:

Optional temp jitter is then added, followed by boundary handling and optional attract pull toward the nearest encoded source patch.

Rendering modes

Decoder

This is the default and conceptually central mode. The final normalised trajectory is mapped back through the robust scaling and inverse PCA rotation into VAE latent coordinates, then decoded into a 64-band × 8-cell log-magnitude representation.

This is a latent spectral-envelope decoder, not a neural waveform decoder.

Nearest event

At every output frame, the nearest encoded source patch is selected in latent space. Its complex STFT content is read through the source-time voice system. This mode reveals the corpus-navigation behavior of the trajectory most directly.

Barycentric

The K nearest source patches are combined according to inverse-distance relationships and converted into equal-power target coefficients before complex STFT mixing. The visualisation reports the actual targeted amplitude coefficients used by this equal-power mix.

Source-time continuity

A separate read pointer measures time since the last retrigger and advances according to rate. A selected donor can therefore continue reading forward through the original source instead of repeatedly looping the same short patch. retrig resets this pointer.

Visualisation

When Draw visualisation is enabled, Latent Pbind creates a multi-panel Praat Picture display designed to explain the entire compositional chain.

Technical behavior

• Processes exactly one selected Sound and never modifies the original.
• Multichannel input is converted to a temporary mono analysis/rendering copy.
• STFT uses a Hann window and 75% overlap; FFT size depends on sample rate.
• Analysis uses 64 triangular log-frequency bands from approximately 40 Hz to the lower of 16 kHz or 95% of Nyquist.
• Low-energy patches more than 60 dB below the strongest patch are excluded from latent training.
• The beta-VAE is implemented directly in NumPy with analytic gradients and Adam.
• Each Pbind key uses an independent seeded RNG derived from the global seed and key name.
• Latent coordinates are PCA-rotated, sign-oriented by spectral-centroid correlation, and robustly scaled by the 2nd...98th percentile range.
• Cluster identities use k-means++ and are ordered along z1.
• Soft-attract navigation uses the geometry of the actual encoded cloud rather than only the rectangular -1...+1 box.
• Decoder output is level-calibrated against the training data and constrained by corpus-derived band and total-energy ceilings.
• Source donor changes use equal-power crossfades.
• Invalid Pbind syntax is parsed before model training, avoiding a wasted training run.
• Per-run temporary filenames include a unique session tag so concurrent Praat runs do not share files.
• All stochastic model and pattern behavior is reproducible from the reported seed as far as the CPU implementation permits.
• The Python engine reports typed errors through a status file and never silently falls back to unrelated behavior.

Requirements & installation

Limitations

• Offline only: the trajectory is composed first and rendered afterward. This is not a real-time latent performance instrument.
• Mono output: multichannel source material is mixed to mono for this tool.
• On-the-fly training: every run trains a model, so processing is slower than purely native Praat transformations.
• Latent axes are source-dependent: z1...zd are learned from the selected Sound and do not carry fixed semantic meanings across different source files.
• Axis orientation is diagnostic: the sign convention points toward increasing spectral centroid only where the reported correlation is meaningfully non-zero.
• Cubic overshoot: cubic interpolation can leave the event-value range before boundary processing.
• Out-of-distribution decoding: large excursions can enter regions unsupported by the training cloud. Distance and level-limiting diagnostics are reported for this reason.
• Source dependence remains: Nearest and Barycentric modes directly reuse source patches; Decoder with Source fine structure also borrows local spectral detail from the source.
• Finite shared-clock streams terminate the composition: this follows the tool's Pbind semantics and should be considered when combining finite and infinite patterns.

Outputs

The script creates a new mono Sound named:

<original-name>_LatentPbind

The original Sound remains unchanged. Output duration is determined by the requested duration, the input duration when Output duration = 0, and any finite pattern that terminates the shared event stream earlier.

Applications

Latent counterpoint

Use case: treat latent coordinates as independent compositional voices with different clocks, cycles, stochastic laws, and interpolation.

Starting point: Radical latent counterpoint.

Polymetric timbral trajectories

Use case: run z dimensions on patterns of different cycle lengths so their combined coordinate never repeats at the period of any single stream.

Starting point: Polymetric coordinates.

Algorithmic identity travel

Use case: use cluster as a macro-formal identity sequence while faster z streams explore local regions around each identity.

Starting point: Cluster pilgrimage.

Structured stochastic exploration

Use case: combine Brownian walks, random holds, weighted choices, and smooth interpolation to produce repeatable but non-literal paths through a learned source space.

Starting point: Brownian latent walk.

Chaotic latent motion

Use case: use logistic-map streams and smoothing to create nonlinear trajectories with deterministic structure but complex local behavior.

Starting point: Chaotic orbit.

Compare latent rendering strategies

Use case: keep one Pbind trajectory and compare Decoder, Nearest, and Barycentric rendering. This separates the compositional path from the synthesis mechanism used to sonify it.