HYPONOIA

(How the Works Above Were Created)

Hyponoia

Hyponoia is an adaptive human–AI system for electroacoustic composition, biointeractive performance, and embodied musical co-creation. It brings together physiological sensing, generative sound organisation, spectromorphological analysis, computational evaluation, human listening, and persistent learning within a single evolving compositional environment. The system has been developed through artistic research in electroacoustic composition, biomusic, physiological computing, and human–AI musical interaction. Rather than functioning as a conventional text-to-music generator, Hyponoia works from a deliberately constructed sonic memory: a curated collection of electroacoustic materials that are segmented, analysed, represented computationally, and subsequently reorganised through generative processes.

 

Sonic Memory and Training

Hyponoia's generative behaviour is informed by a shared training baseline derived from authorised electroacoustic works and curated compositional material. The system analyses this material as individual sound objects rather than treating complete works as indivisible recordings.

Each sound object is described through a combination of acoustic, musical, and spectromorphological characteristics. These include features related to:

• energy and dynamic behaviour;

• spectral brightness and density;

• attack and transient characteristics;

• duration and temporal behaviour;

• pitch and harmonicity;

• resonance;

• noise content;

• richness and musicality;

• textural and ambient potential;

• novelty;

• gesture and morphological development.

 

The resulting representations allow sounds to be situated within a multidimensional musical space rather than classified simply by source identity.

Hyponoia is therefore trained not only to recognise similarities between sounds, but to develop a compositional understanding of how different types of material can function and relate to one another.

Sound objects may take provisional compositional roles such as gesture, texture, resonance, noise, and impact. These categories describe what material can do within a composition rather than what physical source produced it.

 

Spectromorphological Learning

A central development of the current version of Hyponoia is its spectromorphological listening and feedback model.

The system has been calibrated using structured human observations of electroacoustic material, allowing spectromorphological characteristics to become part of its generative decision-making process. The current public Preview incorporates a privacy-safe calibration derived from 27 effective spectromorphological observations, alongside a shared artist-learning baseline derived from 15 authorised electroacoustic works.

Spectromorphology does not operate as a rigid classification system. Instead, it becomes one influence among several that shape the behaviour of the generator.

The system learns relationships concerning qualities such as the evolution of gestures, spectral behaviour, continuity, density, transitions, layering, morphological development, and the interaction between simultaneous materials.

In the current implementation, spectromorphological influence is deliberately bounded rather than absolute. This allows the system to preserve variation and unpredictability instead of forcing every generated composition toward a fixed morphological ideal.

 

Generative Composition

During generation, Hyponoia does not reproduce or reconstruct its training works.

Instead, it draws from its sonic memory and constructs new electroacoustic soundscapes through constrained probabilistic selection and transformation.

The generator considers multiple interacting factors, including:

• the current compositional state;

• learned relationships between sound objects;

• spectromorphological compatibility;

• harmonic and spectral relationships;

• previous selections;

• continuity between events;

• accumulated human preferences;

• compositional roles;

• repetition history;

• novelty and exploration;

• temporal and energetic development.

 

The result is therefore neither predetermined nor completely random.

Hyponoia operates between memory and exploration: previously learned relationships affect future decisions, while controlled probabilistic processes allow unfamiliar combinations and unexpected sonic relationships to emerge.

The system currently supports different compositional conditions — including D1, D3, and D5 — corresponding to different levels of density, activity, development, and energetic behaviour. Generated compositions can be exported both as complete masters and as aligned stems, making the system suitable for further electroacoustic composition, spatialisation, live performance, and artistic research.

 

Human Listening and Adaptive Feedback

Human listening remains central to the architecture.

After a composition has been generated, the artist can listen directly within Hyponoia, view its waveform, navigate through the composition, and attach feedback to precise temporal locations.

This feedback can describe what happens at a particular moment in the work and how its sonic behaviour might develop differently.

Rather than treating human comments as unrestricted commands, Hyponoia converts feedback into bounded compositional advice.

The feedback can influence the next render of the corresponding compositional state while preserving the underlying identity and organisation of the selected musical material. It may alter the expression or relationship of existing events without arbitrarily replacing the entire composition.

This distinction is important: the objective is not to make the system obedient to every instruction, but to establish an iterative relationship between machine generation and embodied musical judgement.

The current learning architecture therefore combines two forms of listening:

computational listening, in which the system analyses measurable characteristics of generated material;

and

human listening, in which an artist evaluates musical continuity, gesture, morphology, balance, development, surprise, and aesthetic effectiveness.

Disagreement between these two forms of evaluation is not considered an error. Instead, it becomes useful information within the learning process.

 

Persistent and Personal Learning

Hyponoia combines a shared compositional baseline with individual local adaptation.

Every artist begins from the same trained system, including its compositional preference model, embeddings, spectromorphological calibration, and shared sonic-learning history.

From that point onward, interaction can become personal.

An artist's feedback gradually influences how Hyponoia behaves on their own computer, creating an individual learning trajectory without changing the common baseline for other users.

The objective is not to optimise the system toward a single universal idea of "good music." Continuous reinforcement of previously successful material would eventually make the system predictable and repetitive.

For this reason, Hyponoia deliberately maintains a tension between learning and exploration.

It remembers previous interactions, but it also allows less familiar materials and relationships to re-enter the compositional field.

 

Biointeractive Performance

Hyponoia was originally developed within a biointeractive performance framework in which physiological information — particularly EEG activity — can influence the behaviour of an electroacoustic environment.

Instead of mapping an individual biosignal directly onto a single musical parameter, physiological activity is interpreted through broader symbolic musical states.

Changes across EEG frequency ranges can therefore modify families of musical behaviour rather than simply controlling volume, pitch, or another isolated parameter.

The body is consequently not treated as a biological MIDI controller.

Physiological variation changes the conditions under which the musical system operates. The resulting sound is returned to the performer through listening, potentially altering attention, gesture, performance behaviour, and subsequent physiological activity.

The relationship becomes recursive:

body → physiological state → computational interpretation → musical behaviour → sound → listening → body

This principle is now being extended into the generative architecture so that physiological state, learned musical memory, spectromorphological behaviour, and adaptive generation can participate within the same compositional ecosystem.

 

An Inspectable Human–AI System

Hyponoia is intentionally designed as an inspectable artistic AI system.

Its compositional process is not hidden entirely inside a single opaque generative model. The principal stages — sound-object analysis, representation, selection, generation, spectromorphological influence, computational evaluation, human evaluation, and adaptive learning — can be examined independently.

This is particularly important within artistic research because the aim is not simply to generate convincing audio output. The project asks how musical agency can emerge across a network of human bodies, listening practices, compositional memory, algorithms, physiological sensing, and adaptive computation.

Hyponoia therefore does not attempt to replace the composer or autonomously imitate a finished compositional style.

It operates instead as an evolving co-compositional environment.

The system remembers.

The artist listens.

The system proposes.

The artist responds.

And each encounter can subtly alter what becomes musically possible in the next one.

​

Listen — Generative Training Progression

Living Ecosystem — D5 | Baseline Training

An early representative render from the adaptive generative sound engine.

https://soundcloud.com/vt8q5dptsdsl/hyponoia_v3_memory_bloom_smoot?si=4d32375ca9d8498cb2729b51f24681ca&utm_source=clipboard&utm_medium=text&utm_campaign=social_sharing

Living Ecosystem — D5 | Intermediate Training

A later stage of adaptive training following iterative human evaluation.

https://soundcloud.com/vt8q5dptsdsl/merit-training-d5?si=6edf614dc5f74503970ee707bd15afb1&utm_source=clipboard&utm_medium=text&utm_campaign=social_sharing

Living Ecosystem — D5+ | Extended Training

An extended training state demonstrating the evolving relationship between sonic memory, exploration, and adaptive evaluation.

https://soundcloud.com/vt8q5dptsdsl/living_ecosystem_d5?si=9a3bf29a7684447e8bdb7eca09e0db83&utm_source=clipboard&utm_medium=text&utm_campaign=social_sharing

​

Code & Open Research

Hyponoia — Max/MSP

Biointeractive live-performance environment and system integration.

https://github.com/penbek7-maker/hyponoia-maxmsp

Muse Microstates

EEG processing and neuro-responsive state-analysis tools associated with the Hyponoia research framework.

https://github.com/penbek7-maker/muse-microstates

 

Hyponoia Generative Composer

https://github.com/penbek7-maker/hyponoia-releases/releases/tag/v2.1.0-preview.2

​

​