English overview · 日本語レポート · Timing visualization
Can audio predict when he becomes shirtless?
Audio was extracted from 23 long-form YouTube DJ/live sets. RMS energy, bass energy, onset strength, spectral centroid, and BPM were measured in 30-second windows. For the 13 sets with a clear visual confirmation, the window immediately before the event was compared with seeded pseudo-event windows from the same set after the visual-label re-audit.
Conclusion
The current data does not support practical prediction from energy or BPM alone. Relative BPM shows a weak signal, while RMS energy is close to chance and the immediate energy rise does not predict the event.
This is not a causal negative result. The visual label is a first clear confirmation, not necessarily the physical undressing moment, and 11 unobserved/inconclusive videos are not valid negative examples.
Event-aligned audio features
Confirmed events are aligned at 0 seconds. Red is within-set RMS energy z-score; blue is within-set BPM z-score. Only the pre-event portion can be used for prediction.
Predictive ranking: event vs pseudo-event
AUC of the preceding window against 215 seeded same-set pseudo-event windows. 0.5 is chance; 1.0 is perfect ranking.
Confirmed event table
| Set | Event (s) | Set position | Pre energy z | Pre BPM | Energy rise z |
|---|
Limits
- BPM uses per-window beat tracking with half/double-tempo normalization.
- Energy and band features are robust z-scores within each set.
- Dallas begins with a shirtless confirmation and has no preceding audio window, so it is not part of the predictive score.
- This exploratory test does not control for track identity, venue, editing, lighting, or personal routine.