Pig vocalizations contain shared acoustic structure for humans and machines, but limited evidence for presumed affective valence

Abstract

Vocalizations are increasingly proposed as indicators of affective state in animal welfare research. Yet many studies assign context-derived affective valence to vocalizations and then classify these using machine learning according to those context-derived labels. This circular dependence makes it unclear whether successful classification reflects affective state itself, broader contextual or acoustic differences, or the interpretive categories imposed by the task. Therefore, we examined human organization of pig vocalizations using free-classification and forced-choice tasks, and compared these patterns with acoustic structure recovered by convolutional neural network models. In a free-classification task, 224 participants sorted 2,192 pig vocalizations into self-defined categories. Next, in two forced-choice tasks, 159 participants recruited in a second wave classified vocalizations using predefined context and valence categories. Free classification revealed reproducible but broad perceptual structure rather than recovery of discrete recording contexts. Participant-generated labels for pig vocalizations were predominantly descriptive and spontaneous valence-related labeling was limited (19.6%) yet primarily negative. Forced-choice classification of recording context was weak (8.0% exact accuracy) and showed only slight agreement with source contexts. Valence judgments were more structured (60.1% exact accuracy), but agreement with the valence categories used to characterize the recording contexts was modest and largely driven by highly aversive situations such as castration, restraint, fighting, and crushing. After excluding pig vocalizations from these contexts, agreement with context-associated valence categories disappeared. Human-derived perceptual structure closely corresponded to convolutional neural network embedding spaces, indicating that human listeners and machine-learning models recovered similar acoustic organization. These findings suggest that pig vocalizations contain robust and recoverable acoustic organization, but that this organization only partially aligns with the contextual and valence frameworks commonly used to interpret it. More broadly, the results highlight a distinction between recovering acoustic structure and establishing its biological meaning, with implications for affective research and animal welfare assessment. Significance StatementAnimal vocalizations are increasingly used to infer affective state and welfare, yet affective labels are often assigned from the recording context rather than independently validated. We show that humans and machine-learning models recover similar acoustic organization in pig vocalizations, especially for highly aversive situations, but that this organization only partially corresponds to the context-derived valence categories used to interpret the vocalizations. These findings show that successful classification of vocalizations based on context-derived presumed valence does not necessarily establish their affective or welfare meaning. Vocal indicators of welfare therefore require validation against behavioral, physiological, and other independent measures rather than reliance on contextual classifications alone.