Fig. 2: Cross-cultural spectro-temporal markers of song vs. speech identified with univariate analyses and machine learning.
From: Spectro-temporal acoustical markers differentiate speech from song across cultures
A Song vs. Speech contrast (two-tailed) in the STM domain across all societies (p < 0.001, FDR corrected in the spectral and temporal modulation domains, n = 369 independent vocalizations). B Heatmap (smoothed) depicting the number of societies showing a significant effect in the clusters identified in (A). Each value reports a numeric count, with larger counts associated with black coloring. C K-means clustering of statistical peaks; dots represent each society. Dark lines illustrate the boundaries of the significant effects presented in (A). D Fieldsite-wise cross-validated support vector machine decoding accuracy (chance level: 50%). The colored dots represent the accuracy for each society (sorted as a function of accuracy with a jet colormap) n = 21 independent societies. E Receiver operating characteristic curve (ROC) for each society (same color code as in (A). Black dashed line represents the chance level. F Normalized feature weights in the modulation power spectrum domain showing features with the largest influence (z-score, average of the 21 classifiers) for the classifier. Dark lines illustrate the boundaries of the significant effects presented in (A).
