AI pipeline quantifies piano composers' harmonic style with 97.9% transcription accuracy

Researchers have built an end-to-end AI pipeline that turns raw piano audio into quantitative stylistic fingerprints. The system first transcribes recordings into musical events using a model certified at F1=0.9791 on the MAESTRO v3.0.0 benchmark—nearly perfect note detection. Then it analyzes harmonic scale-degree distributions through Shannon entropy, asymmetric KL divergence, and Zipfian rank-frequency modeling. Applied to 1,238 works from 15 historical composers (Baroque to early 20th century) plus five neoclassical artists (Richter, Frahm, Glass, Arnalds, Jóhannsson), the pipeline produces interpretable profiles. Historical composers cluster in a narrow entropy band (3.33–3.86 bits), showing that tonal vocabularies are surprisingly similar at the marginal level. Yet KL divergence cleanly recovers known teacher-student lineages—Haydn-Beethoven, Liszt-Rachmaninoff, Schubert-Schumann—with Mendelssohn as an outlier.

The most striking result: neoclassical composers exhibit far more Zipfian regularity in chord transitions (mean R²=0.78 vs 0.46 for historical, N≥10). This gap exceeds variation within either group and aligns with minimalist composition—a smaller, more predictable harmonic vocabulary. The pipeline uses Laplace-smoothed bootstrap 95% confidence intervals for all estimates, ensuring statistical rigor. This technique could extend to other instruments, genres, or even speech analysis, offering a data-driven way to audit artistic style. For AI music generation, it provides a quantitative target: style is not just notes, but the statistical predictability of harmonic choices.