Praat AudioTools · Vector Chain · User Guide

Generative Exploration Map

Turn one recording into a landscape of creative possibilities. Choose a sound-generation engine, audition different results, mark the ones that interest you, and explore further until you find a version to render in Praat.

Author: Shai CohenDepartment of Music: Bar-Ilan UniversityLicense: MITPraat AudioTools repository
2 enginesTwo ways to recompose a recording
24 previewsSuggested starting exploration
3 actionsExpand the map through listening

What does Generative Exploration Map do?

Generative Exploration Map is a listening-guided exploration tool, not a sound generator of its own. You begin with a Sound in Praat and choose a processing engine. The script asks that engine to create many different versions of your recording, using different parameter combinations. It then opens an interactive map in which you can listen to the results and decide which directions to explore.

Each dot on the map is one rendered audio variation. You can hear it, inspect the settings that produced it, mark promising results, and request new variations based on your choices. When you find a result you want to keep, the selected engine renders it from the full recording and returns a new Sound to Praat.

Source SoundSelect a recording in Praat
Choose an engineMarkov or Self-Attention
Explore previewsListen and mark promising points
Grow the mapAround · Between · Other seeds
Render in PraatProduce your chosen Sound
The central idea: the program offers possibilities, but your listening determines what happens next. There is no automatic judgement of which result is musically best.

Choose how the recording will be recomposed

The Processing_engine setting in Praat offers two distinct approaches. The map, auditioning controls, and exploration workflow are the same for both, but the sounds are generated differently.

Engine 1 · Granular and probabilistic

Markov Soundscape Weaver

Divides the recording into short sound grains, learns statistical relationships between them, and uses a Markov process to generate new sequences. Different settings change grain size, the number of states, randomness, overlap, pitch variation, density, and stereo behavior.

What you hear: new soundscapes and textures built from the source's own acoustic material.

Try it for: evolving sound masses, rhythmic fragmentation, unexpected juxtapositions, and transformations of environmental or instrumental recordings.

Engine 2 · Context-sensitive rearrangement

Self-Attention Recomposer

Divides the recording into chunks, describes their acoustic characteristics, and chooses how the chunks follow one another using a context-sensitive similarity process. Different settings control chunk length, context, selection behavior, starting point, repetition, and transitions.

What you hear: new arrangements of recorded gestures, from relatively coherent sequences to more surprising reorganizations.

Try it for: reordering phrases and gestures, exploring continuity versus contrast, and investigating how musical form changes when its parts are rearranged.

These are different kinds of exploration. Markov Soundscape Weaver generates grain-based recompositions. Self-Attention Recomposer rearranges larger source chunks according to relationships among them. The map lets you explore the parameters of either engine; it does not combine the two engines within one variation.

Getting started

  1. Select a Sound in Praat. Choose a recording at least 0.5 seconds long in the Objects window.
  2. Open Generative_Exploration_Map.praat. Set Processing_engine to Markov Soundscape Weaver or Self-Attention Recomposer. The corresponding processing script must be installed.
  3. Choose a preview source. Start with First segment (fast), Preview_duration_s = 5, and Initial_points = 24. These are suggested starting settings rather than a required recipe.
  4. Let Praat generate the initial previews. The script renders a diverse set of configurations. A separate exploration window then opens with one dot per result.
  5. Click dots to listen. The selected preview plays, and the left panel displays its parameters, audio measurements, and rendering information.
  6. Mark interesting dots and explore. Use Mark / unmark (M) to choose points for Around, Between, or Other seeds. New results are added to the existing map.
  7. Choose your final result. Click the point you want, then choose RENDER CURRENT (full length). The processing engine creates a new Sound in Praat.
Selecting is different from marking. Clicking a dot selects it for listening and final rendering. Marking a dot designates it as a starting point for further exploration. Around needs one marked point; Between needs at least two.

Understanding the exploration map

The map helps you navigate many results without listening to them in a fixed order. It provides two ways of arranging the same set of previews:

Sound Space

Dots are positioned using a two-dimensional projection of measured audio characteristics, including spectral, dynamic, onset, and other descriptors. Results that occupy nearby positions have similar measured profiles, although they may still sound different to you.

Parameter Space

Choose the parameters displayed on the horizontal and vertical axes. This view shows where particular settings lie and how parameter changes relate to the sounds you hear. The available parameters depend on the chosen engine.

Exploration map: listening and selectionSchematic illustration, not actual analysis data
■ Generative Exploration Map
Engine: Markov Soundscape Weaver
24 preview points
[View]
Sound SpaceParameter Space
[Colour]
Origin / Round / Parameter
[Point]
Click a dot to hear its sound and inspect its settings.
[Mark]
Mark / unmark (M)
marked: #7
[Explore marked points]
new points: 12
AroundBetweenOther seeds
[Finish]
RENDER CURRENT
Manage Cache… · Close · PANIC
PC1: sound-space axisPC2: sound-space axis MarkedCurrent

Every dot represents an audio preview. A green outline indicates the currently selected point; a white ring identifies a point marked for further exploration. Positions are calculated from audio measurements, not drawn by the user.

Colour and history: you can colour points by how they were created, by exploration round, by a parameter, or, where relevant, by whether the result is deterministic or stochastic. Optional lines show which marked points led to the newest variations. The map preserves earlier points as you continue exploring.

Three ways to explore further

AroundMark one or more promising results. Generate new configurations with parameter settings close to each marked point. Repeat this action to explore a narrower region.
BetweenMark at least two results. Explore intermediate parameter combinations between them. This is not audio crossfading or waveform morphing.
Other seedsKeep the same effective parameters and try different random realizations. This is useful when chance is part of the engine's behavior.

Use the new points control to decide how many additional variations to request. When you choose an action, the map temporarily closes while Praat renders the new previews; it then reopens with your previous results and marks.

For Self-Attention Recomposer, some configurations are deterministic. If selection is Greedy and the start is not random, changing the random seed cannot change the result, so Other seeds is unavailable for those points.

Important distinction: Sound Space organizes the finished previews by acoustic descriptors, but Around and Between explore the engines' parameter settings. Two nearby dots are not necessarily produced by nearby parameter values.

Praat dialog settings

SettingDefaultWhat it means
Processing_engineMarkov Soundscape WeaverChoose which synthesis/recomposition script will generate the previews and final Sound.
Preview_sourceFirst segment (fast)Analyze a short source segment for quicker experimentation, or analyze the whole recording.
Segment_start_s0Where the analyzed segment starts when using the first preview-source option.
Preview_duration_s5Target duration of short audio previews; in segment mode, also sets the analyzed segment length.
Initial_points24How many different configurations to try when creating a new map.
Base_seed1Initial random seed, used for repeatable exploration.
Final_duration_s00 uses the engine's natural output length. A positive value requests a custom duration.
Discard_cached_resultsOffDiscard the matching saved exploration and start again. Leave off to continue an existing map.
Draw_visualizationOnRequest the engine's visualization for the final Sound.
Play_resultOnAutomatically play the final Sound in Praat.
Which preview source should you choose? First segment (fast) is a good way to discover promising settings quickly. Whole sound includes material from the full source during preview analysis, but can take longer. In either mode, the final result is rendered afresh from the complete recording, so it may differ from the short preview.

Rendering your chosen result

RENDER CURRENT (full length) sends the settings of the dot you are currently listening to back to the selected engine. Praat renders a new Sound based on the whole source recording. The original Sound remains unchanged.

Mark / unmark

Choose one or more results as starting points for further exploration. Marking does not render the final Sound.

RENDER CURRENT

Render the currently selected result. It does not need to be marked.

Output names identify the source, engine and selected point, for example:

MyRecording_GEM_MSW_p007
MyRecording_GEM_SAR_p012

With Markov Soundscape Weaver, a positive Final_duration_s requests that duration directly. With Self-Attention Recomposer, the requested duration is converted into a number of chunks; crossfades and chunk boundaries can make the actual result shorter or otherwise different from the requested time. A value of 0 uses the full source-derived chunk pool.

A preview is a listening guide, not a promise of an identical full-length sound. Short-segment previews are generated from less source material. In whole-sound mode, the Self-Attention engine also auditions shorter arrangements of the full chunk pool. The final render uses the complete source and may have a different sequence and duration.

Saved maps and continuing an exploration

Your preview recordings, selected points, marks, and exploration history are saved automatically. Open the same source with the same processing engine and compatible preview settings to return to an existing map without regenerating completed variations.

Each processing engine has its own map. Switching from Markov Soundscape Weaver to Self-Attention Recomposer starts or opens a separate exploration of that recording; the two sets of results are not mixed.

Use Manage Cache… in the map window to inspect storage, open the cache folder, clear the current exploration, or clear saved maps after confirmation. Deleting a map removes its saved previews and exploration history, not your source Sound or separately exported audio.

Remember: Initial_points applies to a newly created map. To add more points to an existing map, use Around, Between, or Other seeds.

Requirements and practical notes

Required scripts: Generative_Exploration_Map.praat, its Python companion generative_exploration_map.py, and the processing script for your chosen engine: Markov_Soundscape_Weaver.praat (or the supported alternative filename Neural_Markov_Soundscape_Weaver.praat), or Self_Attention_Recomposer.praat.

The tool requires Praat 6.4.02 or later and Python with these packages:

pip install numpy librosa soundfile sounddevice scikit-learn pyqtgraph PySide6

The tool checks that the available processing script has a compatible parameter interface before rendering. Preview playback also requires MP3 decoding support in the installed audio libraries. For Self-Attention Recomposer, use a current official Praat build if your platform's packaged build has problems with MelSpectrogram analysis.

ProblemWhat to do
Exploration buttons are inactiveSelect a dot and mark it. Between needs two or more marks; Other seeds needs at least one stochastic point.
Map closes after requesting more pointsNormal behavior: Praat is rendering the new batch and will reopen the map.
Preview differs from full renderExpected, particularly when analyzing only a source segment. Try Whole sound to use the complete source during preview analysis.
Previously saved results do not appearCheck the selected engine, source audio, preview settings, and Discard_cached_results.
Clicking a dot produces no soundCheck the audio device and audition gain in the map; if necessary, check MP3 decoding support.
Suggested first exploration: use a short experimental sound, start with 24 previews, listen to a few contrasting points, mark one, and choose Around for 12 more variations. Render the result you like only when you are ready to hear it developed from the full recording.