Harmony, feeling, evidence.
Cadence’s affect model is an experiment. Its purpose is to help you find and compare musical possibilities. A chord pattern cannot identify how a particular listener will feel.
What is connected now?
The curated Rock Corpus excerpts supply actual within-library frequencies. A versioned importer can add external chord-shape counts and verified song-affect observations. Imported evidence and listening ratings stay in the owner review queue until approved. The workbench reports when a pattern has no imported observations; it never substitutes a published dataset’s advertised size for an actual match count.
Large-dataset import is pending. Chordonomicon and MuSe have not been imported. McGill validation and numeric chord-level calibration are also not complete. Check the workbench’s evidence panel for the current, subsequently approved coverage.
What the research contributes
| Source | Useful contribution | Boundary |
|---|---|---|
| Chordonomicon | Large-scale chord sequences, section markers, genre and release metadata. | Current dataset card: CC BY-NC 4.0; roughly 680k rows. The 2024 paper reports 667,858 curated tracks. Pin the file version and checksum. Published fields do not provide a verified tonic; transposition scripts do not solve key identification. |
| Kantarelis et al. (2024) | Describes Chordonomicon’s curation and representations. | User charts can simplify harmony. Frequency describes the imported chart corpus, not all music or compositional originality. |
| McGill Billboard | Expert harmonic annotations suitable for independent validation. | Keep validation songs separate from training and tuning. Release 2.0 has 890 annotated chart slots covering 740 distinct songs and is CC0. No McGill validation result is claimed here. |
| Lahdelma & Eerola (2016) | 269 participants evaluated 14 chord stimuli, including inversions, with piano and strings. Minor triads and some sevenths were associated with nostalgia/longing. | Single-chord ratings are not progression probabilities. We have not imported their raw ratings or fitted numeric offsets; the model’s coefficients are not attributed to this study. |
| Akiki & Burghardt — MuSe | Song-level valence, arousal and dominance derived from Last.fm tags. | The linked early release describes 90,408 songs; the later journal dataset reports 90,001. Tags are not distributed with this release, and some Spotify matches are tentative. Song joins need verification before aggregation. |
| Makris, Agres & Herremans (2021) | Valence-conditioned lead-sheet generation, evaluated through listening. | Evidence that valence control can work in that system; it does not validate Cadence’s coefficients or generalize to every emotion. |
| Micallef Grimaud & Eerola (2022) | In the interactive experiment, mode and tempo discriminated emotions most strongly; articulation and pitch followed closely. | Dynamics and brightness had smaller, nonzero effects. These experimental results do not supply universal weights for this player. Cadence’s “feel” also bundles rhythmic and articulation changes. |
The evidence pipeline
- Count comparable patterns. Preserve chord quality and bass. For charts without a verified tonic, count transposition-invariant chord shapes, including rotations, without inventing a Roman-numeral key. Only use complete contiguous windows within section boundaries; count each source song once per shape.
- Join conservatively. Require a verified recording identifier or an explicitly reviewed artist/title match. Reissues, covers, remixes and tentative identifiers can cause false matches. Missing joins remain missing.
- Aggregate distributions. Count unique songs, retain means, variances and below-neutral proportions. A whole-song tag is weak evidence for an excerpt. Chorus, verse and bridge weights from the addendum remain untested proposals, so v1 gives each song one equal vote.
- Separate a model from measurements. The prior is an AUWEN heuristic. A 50-song regularizer blends it with imported song estimates. Below 20 matched songs its contribution exceeds half; at 500 it is about 9%. This is transparent shrinkage, not a calibrated Bayesian model.
- Compare playback settings. Tempo and feel adjust modeled arousal; mode and chord quality affect the cold-start estimate. The model does not know the original recordings’ tempo, instrumentation or listeners’ backgrounds.
- Review feedback. One account has one current rating per exact playback context. Owner-approved listener ratings appear separately from song-tag estimates. Free feeling words remain qualitative until a licensed lexicon and mapping have been validated.
Search and rarity
Feeling coordinates are editable-in-code design targets, not universal definitions. Search keeps era, cadence, arc and rarity restrictions. If emotional tags have no exact match, it relaxes only those tags and sorts nearby candidates by modeled valence/arousal. The interface says when this happens. Factual filters can still legitimately produce no results.
Rarity is 100 × (1 − matching source songs / eligible source songs), rounded for display. The denominator is always shown. A relative-to-most-common-pattern score would measure a different thing. Rotations are pooled for frequency only; affect retains tonic, mode, order and bass to preserve harmonic function.
What stays separate
The doctrine of the affections and Tonartencharakteristik remain historical annotations. They are never scoring inputs. Shared progression links retain their v1 format. This pipeline does not fetch or retain lyrics, melodies or recordings.
Sources checked 15 September 2026. Model version: exploratory-1. The next empirical gate is a pinned corpus import, verified MuSe overlap, and song-disjoint evaluation against expert annotations and listening ratings.