Fig. 8.

Classifier generalization across call types and acoustic features supporting classification. a Performance of the regularized LDA classifier at discriminating between pairs of vocalizers when trained either with a distinct set of renditions from the Same call type (S) or with the renditions of all Other four call types mixed together (O). If identity classification rules are shared by all call types, performance is expected to remain unchanged between S and O conditions. Performance drops drastically between Same and Other (χ²₁ = 4710, p < 10⁻¹⁶) but remains above chance level for 5/8 call types. b Performance of the regularize LDA identity classifier on pair-wise sets of call types with the training call type indicated in columns and the tested category indicated in rows. Red stars indicate PCC significantly above 50% (binomial test, p < 0.05). The overall identity classification performance for these significant 24/56 cases was relatively poor (59 vs. 80% for the diagonal). c Classifier performance (PCC) for categorizing pairs of vocalizers averaged across all call types (Identity, purple) or classifying call types irrespective of the vocalizers (Call Type, green). The classifiers were trained on four different acoustic feature spaces: All, all 18 acoustic parameters (see Methods); Spect, 8 spectral parameters only; Temp, 5 temporal parameters only; Fund, 5 fundamental parameters only. For both Identity and Call Type classification there was a significant effect of the feature space but the effect size was much larger for Call Type than for Identity (Identity effect size: 28% increase in odds for the spectral features relative to fundamental or temporal; LRT on GLME: χ²₂ = 853.76, p = 4.054496×10⁻¹⁸⁶; CallType effect size: 64% increase in odds for spectral features relative to fundamental and 118% increase in odds for spectral relative to temporal; LRT on GLME: χ²₂ = 13.894, p = 0.0009614; for call type analysis, acoustical features are averaged per bird across all renditions for a particular call type yielding smaller sample size). In a and c, error bars indicate 95% confidence intervals