Below-1 kHz resonances go unpunished by the prediction model despite lowering actual preference scores

Evidence

Figure 15.14 shows a scatterplot of the predicted vs. measured preference ratings for the 31 headphones with the 95% confidence intervals. The Pearson correlation coefficient (r = 0.86) indicates good agreement between the predicted vs. measured preference ratings with a root mean square error (RMSE) of 6.7 preference rating on a 100-point scale. Two of the headphones (HP6 and HP30) were outliers and fall outside the confidence limits: the predicted ratings were much higher than the ratings they received. The measurements of the headphones suggest the lower observed ratings could be attributed to a large resonance above 1 kHz that was underestimated given by the fact that the explanatory variables SD and AS don’t consider the Q or frequency of the deviation, only their magnitude (see Olive et al. (2018a, appendix 2)). A future version might give more weight to deviations associated with resonances. If HP6 and HP30 are excluded from the model, the goodness of fit improves to r = 0.90. — Toole, Floyd E, Sound reproduction- the acoustics and…, 1813

Categorization

• XREF: The resonance-Q blind spot parallels the loudspeaker finding that narrow resonances matter less than broadband spectral balance — the Q-weighting question recurs across this book’s model development.
• SEED: The model’s failure on HP6 and HP30 is the interesting gap: a purely magnitude-based (no Q, no frequency) model breaks specifically on above-1 kHz resonances — an angle on where simple preference models hit their limits.