Fig 1.
Tension model generation flow.
Displayed is a schematic representation of the model including the feature extraction, the tension prediction involving an attentional and a memory window, as well as the global integration of the feature trends. A: The features are extracted automatically using music information retrieval methods in Python. B: To predict tension, the feature time series are divided into sliding attentional windows (Step 1) and the slope of every feature is extracted in each attentional window (Step 2). Each slope is then integrated with the directly preceding slope using memory windows (Step 3). If the direction of the slope in the memory window matches the direction of the slope in the attentional window, the slope is amplified by β = 5. C: Tension is predicted from the weighted and summed smoothed feature trends.
Fig 2.
Correlations between the tension ratings and the tension prediction for the test pieces in each cross-validation fold.
Displayed are the time-lagged Spearman correlations between the predicted tension and the mean tension ratings. The square dots in dark green indicate the time-lagged Spearman correlation values between the mean tension ratings for the held-out test piece and the tension prediction from the respective validation fold for the time scale model. In light green, the correlations between the tension ratings and the model predictions for the weighted model are plotted. The error bars represent the 95% confidence interval of the Spearman correlations. ** p < .01, * p < .05.
Fig 3.
Displayed is the variation of the model parameters, i.e., weights and window sizes, across the 38 cross-validation folds. A: The distribution of the weights for the time scale model is plotted in dark green and the distribution of the weights in the weight model is displayed in bright green. Every dot stands for one cross-validation fold. B: Plotted is the distribution of the feature window sizes for the time scale model. The attentional window sizes are plotted in dark blue and the memory window sizes are plotted in bright blue. C: Plotted is the distribution of the global window sizes for the weighted model. The attentional window sizes are plotted in dark blue and the memory window sizes are plotted in bright blue.
Fig 4.
Comparison between the mean tension ratings and the tension predictions from the optimal model configurations.
Displayed are the tension predictions and the mean tension ratings for three example pieces taken from our sample. The mean tension ratings are displayed in black. Predictions from the time scale model are plotted in dark green and predictions from the weighted model are plotted in bright green. The error bands show the standard error around the mean of the tension ratings. The mean tension ratings have been shifted by 4.5 seconds to account for the delay in reporting behavioral tension and facilitate the visual evaluation of the overlap between the curves.