The Arousal Index

An interpretable, calibrated, cross-dataset model of musical arousal.

This repository accompanies the paper "The Arousal Index: An Interpretable, Calibrated Model of Musical Arousal that Generalizes Across Datasets" — submitted to ICASSP 2027 (under review). The submitted PDF is in paper/. arXiv preprint and archival DOI: forthcoming — links will be added here as they go live. It contains a compact, fully interpretable model that predicts the arousal (felt activation/energy) a listener experiences from a piece of music, on a calibrated 0–11 scale, from three named acoustic features — loudness, noisiness, and a tonality index — each summarized by its mean and variability (six descriptors), selected from six extracted features by a collinearity analysis. The headline result: this decorrelated three-feature model transfers across datasets substantially better (Spearman 0.77) than a twelve-descriptor model built from the same six underlying features (0.60) — and spline and gradient-boosted variants of the same six descriptors fit better in-corpus while transferring worse still (0.49, 0.40). A 6,373-functional openSMILE black-box (same boosted learner, features extracted identically on both corpora) pushes within-corpus R² to 0.63 — the best of any model tested — while transferring at 0.43: the largest capacity, the largest accuracy–transfer gap (bootstrap Δρ vs. linear +0.34, 95% CI [+0.28, +0.40]). In the reverse direction (train PMEmo → test DEAM) capacity again never helps: spline ties linear (ρ ≈ 0.59) and boosting drops to 0.52. Across both directions, added capacity never improved transfer — and severely harmed it when training on the broader corpus. Resolving collinearity and restraining capacity improved both interpretability and cross-corpus generalization.

A note on terminology. This repository is the umbrella for a research program targeting musical intensity. Its first model (this work) honestly measures arousal as operationalized by human listener ratings — a related but distinct construct that includes production factors such as loudness and mastering. Work on intensity proper (which isolates sonic aggression from production) is future work. Throughout, "arousal" means what the model measures. A second terminology note used throughout the code and paper: a feature is a named frame-level acoustic quantity; a descriptor is its per-clip mean or standard deviation.

The interpretable model is the least accurate within a single dataset and generalizes far better than every higher-capacity alternative — added capacity partly memorizes dataset-specific structure, which the transfer test exposes. Selecting a model by within-dataset accuracy would pick the black-box and forfeit 0.34 of transfer correlation (0.77 → 0.43). Reverse direction (train PMEmo → test DEAM): linear ρ = 0.59, spline 0.59, boosted 0.52 — capacity never helps in either direction; the severe penalty is specific to training on the heterogeneous corpus.

├── paper/farrell-arousal-index-icassp2027.pdf   # the submitted paper
├── arousal-index.ipynb          # the complete pipeline, runs top-to-bottom
├── deam_librosa_features.csv    # pre-extracted librosa features, DEAM (1802 songs)
├── pmemo_librosa_features.csv   # pre-extracted librosa features, PMEmo (767 songs)
├── arousal_index_curves.png     # Figure: interpretable feature-effect curves
├── refs.bib                     # verified bibliography for the paper
├── requirements.lock            # pinned dependency versions
└── README.md

The two openSMILE feature tables used by the black-box reference (deam_opensmile.csv, pmemo_opensmile.csv, ~300 MB combined) exceed GitHub's file-size limits and are published separately as a Kaggle dataset: https://www.kaggle.com/datasets/connerofarrell/deam-and-pmemo-opensmile. The final notebook cell finds them automatically when that dataset is attached (or the CSVs are placed in the working directory), and otherwise re-extracts them from the source audio (~2.5 h; requires internet for pip install opensmile).

pip install -r requirements.lock

(Python 3.11+. The pinned versions are those used to produce the reported numbers. opensmile, used only by the final black-box cell, is installed at runtime by that cell.)

This repo ships the pre-extracted librosa features (the two smaller .csv files), so you can reproduce every core number without downloading any audio. Just run the notebook — it detects the CSVs and uses them (the "fast path").

To re-extract features from scratch, or to use the audio-based scoring tool, download the two public source datasets (both free for non-commercial research; not redistributed here due to size and licensing):

• DEAM (MediaEval Emotion in Music): https://www.kaggle.com/datasets/imsparsh/deam-mediaeval-dataset-emotional-analysis-in-music
• PMEmo (2018): https://www.kaggle.com/datasets/adityaraghuvanshi999/pmemo-original

The feature extraction recipe is defined in extract_six() near the top of the notebook — it is the single source of truth for how the librosa CSVs were produced. The openSMILE extraction recipe lives in the final cell and uses the identical 30 s / 22.05 kHz front end.

Open arousal-index.ipynb and run all cells top to bottom. On the fast path, everything through the ablation study completes in well under a minute and reproduces: within-dataset R² = 0.29, cross-dataset transfer Spearman = 0.77, the reverse-direction results, the bootstrap confidence intervals, the anchor ladder, and the feature-effect figure. The final cell (the black-box reference) additionally reproduces R² = 0.63 / ρ = 0.43 — fast with the openSMILE CSVs available, ~2.5 h if re-extracting.

After running the notebook, the score_arousal() function scores any audio file 0–11:

score_arousal('your_song.mp3')   # -> e.g. 6.5

Note: the model is trained on DEAM (predominantly folk/acoustic/light-rock) and will underestimate genuinely extreme music (e.g., metal), which lies outside its training range — a known limitation discussed in the paper and motivating future work with wider-range stimuli.

Six interpretable acoustic features — loudness (RMS), spectral brightness (centroid), attack density (onset strength / spectral flux), noisiness (zero-crossing rate), spectral flatness, and a tonality index (negative log-flatness; monotone-related to flatness at the frame level) — each summarized by its mean and standard deviation across 30-second clips (twelve descriptors). Brightness, ZCR, flatness, and tonality form one highly inter-correlated spectral group (|r| = 0.62–0.86 on both corpora), so the interpretable model keeps the group's two least-correlated members (ZCR and tonality) plus loudness — three features, six descriptors. An ablation adding onset strength raised within-corpus fit (R² 0.29 → 0.35) but lowered transfer (ρ 0.77 → 0.65) — its correlation with arousal itself reverses sign between corpora (+0.40 DEAM vs −0.23 PMEmo) — so it is excluded. A standardized Ridge regression maps the six descriptors to a 0–11 arousal scale, calibrated with isotonic regression fitted on out-of-fold predictions. Evaluated within-dataset by 5-fold cross-validation and across datasets by training on DEAM and testing, unchanged, on PMEmo (features extracted identically on both), plus the reverse direction as a robustness check. The black-box reference applies the same gradient-boosted learner (200 trees, depth 3) to the 6,373 openSMILE ComParE 2016 functionals, extracted with the same front end on both corpora.

DEAM and PMEmo are the property of their respective creators and are used here for non-commercial research under their stated licenses; see the linked dataset pages. The pre-extracted feature tables in and linked from this repository (librosa and openSMILE) are derived summary statistics released to support reproducibility.

Please cite the original datasets in any derivative work:

• DEAM: Aljanaki, A., Yang, Y.-H., & Soleymani, M. (2017). Developing a benchmark for emotional analysis of music. PLOS ONE.
• PMEmo: Zhang, K., Zhang, H., Li, S., Yang, C., & Sun, L. (2018). The PMEmo dataset for music emotion recognition. ICMR.

Farrell, C. O. (2026). The Arousal Index: An Interpretable, Calibrated Model of Musical Arousal that Generalizes Across Datasets. Submitted to ICASSP 2027.

@misc{farrell2026arousal,
  author = {Farrell, Conner O.},
  title  = {The Arousal Index: An Interpretable, Calibrated Model of Musical Arousal that Generalizes Across Datasets},
  year   = {2026},
  note   = {Submitted to ICASSP 2027. Code: https://github.com/ConnerFarrell/intensity-index}
}

The arXiv ID and Zenodo DOI will be appended here once live.

Code in this repository: MIT (see LICENSE). Derived feature data: released for non-commercial research consistent with the source datasets' licenses.