An Experiment with AI Music Generation

I experimented with AI music generation by providing structural prompts and listening to the result. The unedited track below is a sample from that experiment.

A listening example can show what a particular output sounds like. It cannot establish how the model was trained, whether its outputs are consistently original, or how reliable it is across prompts.

Questions for a more systematic evaluation

I would separate listening quality from provenance. An evaluation could examine whether the piece follows the requested structure, remains coherent over its duration, and contains audible artifacts. A provenance investigation would need separate evidence about the model, its inputs, and any watermarking method.

My doctoral research examines Proof-of-Learning and model watermarking. Applying related ideas to generated audio would require a new method and experiments specific to audio.

Open research direction of the author, not yet published. No audio-watermark benchmark or deployment result is reported here.