📄 animal behavior and cognition

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

This study demonstrates that while humans and machine learning models can reliably identify shared acoustic structures in pig vocalizations, human perceptual judgments of emotional valence only align with presumed affective states in highly aversive contexts, suggesting a critical distinction between recoverable acoustic patterns and their biological interpretation in animal welfare research.

Original authors: Gorssen, W., Sleurs, B., Winters, C.

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gorssen, W., Sleurs, B., Winters, C.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are at a party where you can't see the guests, only hear their voices. You hear a mix of screams, giggles, grunts, and whines. Your job is to sort these voices into groups based on what they sound like, without knowing who is making them or what is happening around them.

This is exactly what the researchers in this paper asked humans to do with pig sounds. They wanted to see if people could naturally figure out if a pig was happy, sad, angry, or in pain just by listening to a recording, and whether computers (AI) could do the same thing.

Here is the story of what they found, broken down simply:

1. The Setup: Two Different Ways to Listen

The researchers ran two different "listening parties" (studies):

The Free-Form Party (Wave 1): People were given a pile of pig sounds and told, "Sort these into whatever groups you want, and give the groups your own names." No rules, no right or wrong answers.
The Multiple-Choice Party (Wave 2): A new group of people was given the same sounds but forced to pick from a specific list of labels, like "Is this pig being castrated?" or "Is this sound positive or negative?"

They also trained a computer (a Convolutional Neural Network) to look at the sound waves like pictures and sort them the same way.

2. The Big Discovery: The "Pain" Signal is Clear, But the Rest is Foggy

The results were surprising and nuanced.

The "Scream" Analogy:
Think of the pig sounds like a spectrum of weather.

The Storms (High-Aversive Situations): When pigs were in terrible situations—like being restrained, fighting, crushed, or having their tails docked (castrated)—their sounds were like thunderstorms. Everyone, whether a farmer, a city dweller, or a computer, agreed immediately: "That sounds terrible!" These sounds were distinct, loud, and clearly negative.
The Drizzle (Other Situations): For everything else—like pigs running around, eating, or being in different types of housing—the sounds were like drizzle or mist. They all sounded somewhat similar. When people tried to sort these, they couldn't agree on which "drizzle" belonged to which "cloud." A sound made while a pig was "running" (presumed happy) sounded very similar to a sound made while it was "exploring" (presumed neutral).

3. What People Actually Said

When people in the "Free-Form Party" sorted the sounds, they didn't mostly use emotional words like "sad" or "happy."

They mostly used descriptive words. They said things like "loud," "scary," "pig," or "squealing."
They only used emotional words (like "pain" or "fear") about 20% of the time, and almost exclusively for the "storm" sounds (the terrible situations).
When they did guess the emotion, they were mostly right about the "storms" being bad, but they were often wrong or confused about the "drizzle."

4. The Computer vs. The Human

The researchers compared how humans sorted the sounds with how the AI sorted them.

The Match: Humans and the AI were very similar. They both heard the same acoustic patterns. If the computer thought two sounds were alike, humans usually thought they were alike too.
The Limit: Just because the computer and humans agreed on the sound structure doesn't mean they agreed on the meaning. The AI could perfectly separate the "storms" from the "drizzle," but it couldn't reliably tell the difference between a "happy run" and a "neutral walk" based on sound alone.

5. The Main Lesson: Hearing the Sound vs. Knowing the Story

The paper's most important point is a distinction between hearing and interpreting.

Hearing the Sound: We (and computers) are great at hearing that a pig is making a specific type of noise. We can tell that a "castration scream" sounds different from a "feeding grunt." The acoustic structure is real and recoverable.
Knowing the Story: We are not great at guessing the pig's internal feelings just from that noise. The paper argues that just because a sound can be sorted into a category (like "negative"), it doesn't prove that the pig is actually feeling "negative" in the way we think.

The Takeaway Metaphor:
Imagine you hear a car engine revving loudly.

Acoustic Structure: You know it's a loud, high-pitched rev. (Humans and AI agree on this).
Interpretation: You might think, "That driver is angry!" or "That driver is racing!" or "That driver is just warming up the engine."
The Paper's Warning: The paper says that for pigs, we are often too quick to say, "That sound means the pig is sad." The study shows that while the "sad" sounds (the storms) are clear, most other pig sounds are a mix of many different feelings that sound too much alike to tell apart without seeing the pig's face or knowing what's happening around it.

In short: Pig sounds have a clear "grammar" that humans and computers can learn, but that grammar doesn't translate perfectly into a simple "Happy vs. Sad" dictionary. We can hear the pig is making noise, but we can't always be sure exactly what that noise means about its feelings without more clues.

Technical Summary: Pig Vocalizations, Acoustic Structure, and Affective Interpretation

Problem Statement
Animal vocalizations are increasingly utilized as indicators of affective state and welfare in research. However, a methodological circularity often exists: studies frequently assign affective valence to vocalizations based on the recording context (e.g., assuming a sound from a castration procedure is "negative") and then use machine learning or human listeners to classify these sounds based on those same context-derived labels. This approach makes it unclear whether successful classification reflects the animal's internal affective state, broader acoustic differences, or merely the interpretive categories imposed by the task. Furthermore, existing literature relies heavily on forced-choice designs with predefined categories, leaving unclear how listeners spontaneously organize vocalizations without an interpretive framework. This study addresses the distinction between recovering acoustic structure and establishing its biological meaning, challenging the assumption that acoustic signals directly map to discrete emotional states.

Methodology
The study employed a two-wave online classification design involving human participants and machine-learning models, utilizing a dataset of 2,192 unique pig vocalizations derived from 18 experimentally defined contexts (Briefer et al., 2022a).

Wave 1 (Free Classification): 224 participants sorted 8,960 presentations of vocalizations into self-defined categories without predefined labels. Participants also provided descriptive labels for their categories. This wave included control stimuli (duplicated castration sounds and a dolphin outlier) to assess attention and consistency.
Wave 2 (Forced Choice): A separate group of 159 participants performed two tasks on a subset of 1,357 vocalizations: (1) classifying sounds into 18 predefined recording contexts (including a "no idea" option) and (2) rating the perceived valence of the sounds on a 5-point scale (very negative to very positive).
Machine Learning Analysis: Convolutional neural network (CNN) models, specifically fine-tuned ResNet50 architectures, were trained on spectrograms of the vocalizations. Two models were developed: one optimized for binary presumed-valence classification and another for 18-context classification.
Comparative Analysis: The study compared human-derived perceptual structures (co-classification matrices) with machine-learned acoustic embedding spaces using Mantel tests. It also analyzed the semantic content of participant-generated labels and tested the robustness of findings by excluding high-intensity aversive contexts (castration, restraint, fighting, crushing).

Key Results

Acoustic Structure vs. Context Recovery: Human free classification revealed a reproducible but broad perceptual structure. Participants did not recover the 18 original recording contexts as discrete categories. Instead, they grouped vocalizations into broad families, with the highest similarity observed among high-intensity aversive contexts (castration, restraint, fighting, crushing).
Labeling and Valence: In free classification, participant-generated labels were predominantly descriptive (states, events, animal-related) rather than emotional. Spontaneous valence-related labeling was limited (19.6%) and primarily negative. When forced-choice valence judgments were made, accuracy was modest (60.1% exact accuracy excluding neutral responses) and driven almost entirely by the high-intensity aversive contexts. Agreement with presumed valence disappeared when these aversive contexts were excluded.
Context Classification: Forced-choice classification of the specific recording context was weak, with an exact accuracy of 8.0% (9.5% excluding "no idea" responses), showing only slight agreement with source contexts ( $\kappa = 0.042$ ).
Human-Machine Convergence: Human-derived perceptual structure closely corresponded to the embedding spaces of the machine-learning models. Both humans and models recovered similar acoustic organization, particularly distinguishing high-intensity aversive sounds. However, this shared structure only partially aligned with the context and valence categories assigned in the source dataset.
Expertise Effects: Professional background (farmers, veterinarians, scientists) influenced the type of labels used (e.g., farmers used more pig-specific animal terms; scientists used more affective terms) but had little influence on the underlying perceptual clustering structure.

Significance and Claims
The paper claims that while pig vocalizations contain robust and recoverable acoustic organization that is accessible to both human listeners and machine-learning models, this organization does not necessarily align with the contextual and valence frameworks commonly used to interpret them.

Distinction between Structure and Meaning: The authors emphasize a critical distinction between recovering acoustic structure and establishing its biological meaning. Successful classification of vocalizations based on context-derived presumed valence does not necessarily establish their affective or welfare meaning.
Limitations of Current Welfare Indicators: The findings suggest that vocal indicators of welfare require validation against behavioral, physiological, and other independent measures rather than reliance on contextual classifications alone. The study highlights that for many contexts (e.g., housing conditions, positive conditioning), vocalizations overlap substantially in acoustic space and human interpretation, failing to support the idea of invariant acoustic indicators for specific affective states.
Role of Interpretation: The results support constructionist perspectives, suggesting that the affective meaning attributed to vocalizations is probabilistic, context-dependent, and partly shaped by the categories and assumptions brought to the task by the observer (human or machine).

In summary, the study demonstrates that humans and machines can reliably detect acoustic patterns in pig vocalizations, particularly for highly aversive situations, but cautions against interpreting these patterns as direct, context-independent readouts of discrete emotional states or welfare conditions without independent validation.