yalishanda/dynamics-needed-lgbm

Dynamics Needed — LightGBM velocity baseline

Version: 0.1.0

Gradient-boosted-tree baseline that predicts per-note velocities ("dynamics") for MIDI drum tracks from tabular note features. Trained on the Expanded Groove MIDI Dataset (E-GMD). Part of the Dynamics Needed thesis project.

Intended use

Given a MIDI drum track with flat/undynamic velocities, predict a "best-fitting" velocity per note to restore human-like dynamics. This is the tabular baseline; see the MDN transformer model for the probabilistic variant.

Training data

E-GMD (Expanded Groove MIDI Dataset), evaluated on the held-out test split.

Metrics (test split)

model	MAE	RMSE
Global-mean	29.668	34.638
Lookup table	21.393	28.380
LightGBM	18.020	24.072

Per-track Pearson (LightGBM): 0.706

Limitations

This is a point (single-value) predictor and a baseline — published for reproducibility and comparison, not as a final production model.

• Flattens dynamics. It regresses toward the conditional mean (std ratio ~0.69), so it cannot reproduce the full velocity distribution / ghost-note tails. Restoring that spread is the job of the probabilistic transformer heads.
• Absolute-loudness generalization gap. On drummers unseen in training, point MAE degrades (~19 → ~28); a player's overall loudness is not inferable from structure alone. Relative dynamics (per-track / within-bar ranking) transfer better than absolute level.
• Known data artifact. E-GMD's multi-kit rendering remaps pads to different voices per kit, which biases some per-voice results; a single-kit rebuild is a pending fix. See the project's docs/methodology/kit-remapping-artifact.md.
• No listening test yet. Numbers here are offline metrics; perceptual A/B validation is future work.

License

Set to mit by default — change to match the thesis's chosen license before publishing.