Harmonizing AI: Rethinking Chord Adaptation Across Music Genres

In the ever-expanding world of music generation, researchers are constantly seeking methods to enhance models' adaptability to various musical genres. The latest exploration centers around a symbolic layer called Harmony, where mathematical pitch relations, acoustic consonance, and musical convention intersect.

Extending the Reach of Music Transformer

Starting with a well-established pop-jazz Music Transformer checkpoint, the question on everyone's mind is: how far can small adaptation interfaces take us in extending this model to many genres? The genres under scrutiny include blues, bossa nova, Bach chorales, country, electronic, folk, funk, gospel, hip-hop, R&B/soul, and rock. The main evaluation pits five adaptation methods against each other, LoRA, IA3, BitFit, prefix tuning, and full fine-tuning, across an extensive 165-cell grid of genre adaptation possibilities.

All five methods showed improvements over the frozen base model held-out chord prediction. But, let's apply some rigor here. The macro gains ranged from +2.89 to +3.61 points. While LoRA and IA3 emerged as top performers, statistical tests couldn't confirm a clear winner. Does this make the whole exercise moot? Not quite.

The Data Conundrum

What they're not telling you: when genres are sub-sampled to a uniform corpus size, IA3 stands its ground, but LoRA's advantage dissipates. This highlights an uncomfortable truth in machine learning, sometimes, your model's performance is more about data availability than pure algorithmic superiority. In fact, the control-token baseline often outperformed the frozen base, suggesting that many improvements might simply arise from lightweight conditioning over a shared harmonic foundation, rather than any specific adapter technique.

Chords Aren't the Whole Story

Additional diagnostics, including rank sweeps and wrong-genre rotations, pointed to a sobering conclusion: while chord-symbol adaptation reliably boosts genre-local harmonic predictions, these symbols alone don't embody the full scope of genre identity. So why should we care? Because this underscores the limitations of current models in capturing the full richness of musical genres, challenging us to think beyond mere chord sequences.

As we push these models further, a fundamental question lingers, can we ever expect AI to fully grasp the essence of a genre that human musicians spend a lifetime perfecting? Color me skeptical, but there's an inherent complexity in music that's tough to encode purely through data-driven approaches.