A listening comparison
Short pieces written as scores by language models and played by sampled instruments, a piano or an orchestra, each set beside what three music-generation models made from a description of the piece or from its tune. Level-matched, with every prompt shown, and presented without verdicts.
Every model take was turned up or down by one plain gain change, with no compression or limiting, so that its loudness matches the composed piece within 0.3 LU. Loudness then does not decide the listen. The gain is on every card; the raw numbers are in each piece's table.
No model take on this page has been judged. The notes under each piece state facts (lengths, levels, what each model was given) and leave the listening to you. All three models were run locally with fixed seeds.
Model takes are 192 kbps MP3 copies of the level-matched files; the composed pieces are their original renders. One player plays at a time, and each piece has a Listen blind button that hides the labels and shuffles the cards.