Overview
The AI GAN Music Generator leverages modern client-side audio technology to execute real-time procedural compositions. By mapping user-controlled latent space configurations directly into sound parameters and visual matrices, the system explores the boundaries between user agency and programmatic algorithms. Built entirely to run inside standard modern browsers, this architecture is free of server-side latency limitations and heavy network loads.
How to Use
- Activate Audio: Press the Audio System: OFF button to enable state synthesis safely complying with standard mobile browser interaction requirements.
- Modify Latent Space Nodes: Drag the Melodic Entropy, Rhythmic Density, and Harmonic Spread sliders to shift the coefficients of the neural generative simulator.
- Adjust Generative Entropy (Temperature): Higher values allow for chaotic structural variations, while lower values direct the calculations into deterministic patterns.
- Trigger Generation: Click Generate Neural Matrix to calculate a 32-step by 88-note sequence. You can also click or drag directly on the interactive Piano Roll grid to customize individual notes.
- Play/Stop: Click the Play Sequence button to loop the playhead across the composition.
Technical Details
Under the hood, this generator executes real-time inference via an emulator representing neural layer processing networks. Linear transformations process the 3-dimensional latent vector combined with speed and scale coefficients. These are passed through a non-linear activation function, and then mapped into a tensor representation of size $32 \text{ steps} \times 88 \text{ notes}$ (representing a full piano keyboard).
To prevent harsh acoustic dissonance, this version routes output signals into a dedicated signal processor containing lowpass filters and spatial reverb. Furthermore, the 88 notes are segmented into three registers: a monophonic bass lane, a polyphonic harmony lane constrained to the root progressions of the scale, and a flowing melody lane. This mirrors the compositional rules used in traditional sequencing software.
Future Directions
As client-side tensor calculations advance through GPU-based web pipelines, loading deep learning layers locally will become increasingly streamlined. Future iterations of this simulator will support uploading pre-trained models, exporting generated compositions as raw MIDI datasets, and synchronizing procedural canvas displays with multi-layered frequency visualizations.