The LETHE System: Dataset-Free Self-Referential Audio Generation

Introduction

At the intersection of artificial intelligence and musical composition, the experimental LETHE system is charting a distinctive course: it draws on the architectural ideas of Generative Adversarial Networks (GANs) while completely abandoning any reliance on external datasets. Implemented in SuperCollider, this system achieves continuous sonic evolution through a self-referential mechanism, offering a new technical paradigm for electroacoustic music creation.

Core Principles of the LETHE System

LETHE stands for "Latent-parameter Evolution with Temporal Hierarchical quasi-Equilibrium." At its heart, it is a self-referential sound-forgetting system. Unlike traditional GANs, which require large amounts of training data, LETHE enters a fully closed operational mode after initialization — requiring no external datasets or human supervision.

The system's core is a 3×3 mixing matrix built around two delay lines. The matrix contains 9 coefficients and 2 delay-time parameters, which evolve continuously through the interaction of a five-feature linear discriminator and a stochastic perturbation optimizer. This optimizer resembles a single-sample REINFORCE algorithm, enabling parameter adjustment without gradient information.

Technical Architecture Innovations

LETHE's discriminator design is particularly distinctive: rather than competing against a generator, it compares current energy behavior against an archived snapshot of the initial state, thereby guiding the direction of parameter updates. This design allows sonic characteristics to drift and evolve continuously while maintaining a memory anchor point.

The system supports the independent mixing of three audio sources: looped audio, fixed audio, and live input. This flexibility allows composers to deploy LETHE across different contexts, from purely algorithmic generation to semi-improvised performance.

The research team validated the system's effectiveness through ablation experiments. Results showed that an active generator is essential for parameter evolution — across all 15 ablation control groups, the change in the key parameter $\Delta c_{22}$ was 0.000, confirming the central role of the adaptive mechanism.

The Self-Referential Electroacoustic Tradition

LETHE's design philosophy is rooted in the tradition of self-referential electroacoustic music. In this tradition, composers no longer directly control every note or timbral detail; instead, they construct systems capable of autonomous evolution, then delegate the sonic outcome to this adaptive closed loop.

This compositional approach redefines the composer's role: from note arranger to parameter-space architect. The composer shapes the possibility space of sonic evolution by establishing initial conditions, system rules, and constraint boundaries — while the specific sonic trajectory is determined by the system's internal dynamics.

SuperCollider Implementation

LETHE is implemented in SuperCollider, a programming environment designed for real-time audio synthesis. SuperCollider's powerful signal-processing capabilities and flexible control structures make it an ideal platform for realizing complex adaptive audio systems.

The system's closed-loop design offers several advantages: it requires no large training datasets or significant computational resources; its ability to run independently after initialization suits real-time contexts such as live performance; and its evolutionary path is inherently unpredictable, producing a unique sonic trajectory with each run.

A New Direction for AI Music Creation

The LETHE project reveals an alternative vision for AI music creation: rather than learning from vast corpora of musical data to imitate human composition, it constructs self-organizing systems with their own intrinsic evolutionary logic. This approach aligns more closely with the philosophy of Generative Art, emphasizing process over product, and system over artifact.

From a technical standpoint, LETHE demonstrates that GAN architectural ideas can operate outside the supervised learning paradigm — functioning effectively within a fully closed system. A discriminator need not compare against external ground-truth data; it can guide system evolution through internal consistency checks alone.

This self-referential system provides experimental music creators with a new tool. It preserves the autonomy and unpredictability of algorithmic composition while giving composers sufficient expressive space through the design of the parameter space.

Conclusion

LETHE represents a distinctive branch of AI music system design: rather than seeking to emulate existing musical styles, it explores mechanisms for autonomous sonic evolution. By translating the adversarial logic of GANs into a self-referential parameter optimization framework, LETHE achieves continuous sonic transformation without any external data.

This project reminds us that the application of artificial intelligence in artistic creation need not be confined to the "learn-and-reproduce" model. Building generative systems with intrinsic logic — and partially ceding creative control to algorithmic self-organization — may open up more experimental and genuinely original avenues for artistic expression.