Extended Data Fig. 4: Cross-validation of the metrics.

For each dataset, the metrics were cross-validated to test their generalizability within the dataset. Explained variances (r²) of the human ratings by the cross-validated metrics for each dataset are presented for: the Training correlations (fitted on the N-1 sounds), the Testing correlations (tested on the removed sound), the within correlation between the N*(N-1)/2 metric pairs characterizing the Internal consistency of the fitted metrics in each dataset, and the average correlation (r²) with the metric Refitted on all sounds with those on the N-1 subsets showing the extent to which this metric is different from the cross-validated one. On median, the metrics were cross-validated with r² = 0.51 on the testing sets. For each dataset, they are highly consistent within the N-folds (r²_: Mdn=0.85), and they strongly correlate with the metric fitted on the whole dataset (r²_: Mdn=0.92). For each dataset, the correlation of the refitted metric on whole sounds with those fitted for the cross-validation (last column) are high showing that the metric fitted on whole sounds can be used to perform the analyses.