Pig vocalizations contain shared acoustic structure for humans and machines, but limited correspondence with presumed affective valence
Year: 2026
Type: preprint
Abstract: Pig vocalizations are increasingly used as indicators of affective state and welfare. This view is supported by above-chance human valence judgments and striking machine learning classification performance in prior studies. Yet, in both approaches the affective distinction is already built into the test. Algorithms are trained and evaluated using presumed valence labels assigned to recording contexts, while human listeners are asked directly to rate valence. We therefore asked what structure emerges when listeners are given no predefined categories, and how that structure relates to explicit human judgments and neural network representations. In a free-classification task, 224 participants labeled and sorted 2,192 pig vocalizations. A second group of 159 participants then classified vocalizations using predefined context categories and rated their valence. Humans and neural networks recovered corresponding organization, confirming that pig vocalizations contain reproducible acoustic information. Yet, affect was not the dominant spontaneous interpretation, with only 26.3% of participant-generated labels referring to feelings or emotions. Explicit valence judgments were systematic and agreed moderately with presumed valence, with 60.1% directional agreement. However, this correspondence disappeared when highly aversive situations were excluded, although neural networks still classified presumed valence with 92.0% accuracy after removing these contexts. In a matched permutation analysis, however, neural networks with arbitrary valence reassignments on the context level averaged 79% accuracy, with 5 of 100 exceeding the original grouping. Pig vocalizations therefore contain rich acoustic information that both humans and algorithms can recover, but successful recovery does not by itself establish what that information means for the animal. Progress in vocal indicators of pig affect and welfare will require moving beyond classification performance toward independent validation of the biological meaning assigned to acoustic differences. Significance Statement Animal vocalizations are increasingly used to infer affective state and welfare, yet presumed valence is often assigned at the level of the recording context and then applied to individual vocalizations produced within that context. 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 context-derived valence. More fundamentally, successful classification of presumed affective state is not equivalent to demonstrating that vocalizations provide a valid readout of the animal’s affective state. Their use as welfare indicators therefore requires independent validation against animal-derived evidence of what animals seek or avoid and consequences for their health, rather than inferred from human-assigned emotional categories.
Authors: Wim Gorssen, Ben Sleurs, Carmen Winters
Institution: ETH Zurich
Language: English
Subfield: Animal Science and Zoology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7167800788