Can people hear how a pig feels?

People and artificial intelligence (AI) detect reproducible patterns in pig vocalizations, especially during acute distress. Yet an isolated sound rarely reveals the exact situation or whether a pig feels positive or negative.

Animal-welfare research increasingly explores the use of artificial intelligence (AI) to monitor pig welfare through vocalizations. A common approach is to classify recordings as positive or negative based on the context in which they were obtained. For example, vocalizations recorded during castration are typically considered negative, whereas those recorded during suckling are considered positive. AI systems have been able to distinguish between such categories with considerable success. However, it remains unclear what information these systems actually use: do they detect affective states, general acoustic patterns, or simply the context-based categories used for training?

We examined what people hear when all other contextual information is removed. We used over 2,000 clips from an online pig vocalization database, with known underlying context. Participants saw no video and did not know where or when the vocalizations had been recorded. In a first online experiment, 224 participants grouped 40 clips according to what they heard and named the groups themselves. In the second experiment, 159 participants assigned 25 clips to one of 18 recording contexts or ‘no idea’, then rated the same clips from very negative to very positive. We compared their classifications with the acoustic structure learned by a neural network from spectrograms.

People and AI find similar patterns

People did not group the vocalizations at random. A broad, reproducible acoustic structure emerged, and it was almost the same for people with and without pig expertise. Expertise mainly changed the words used: pig farmers chose more pig-specific descriptions, whereas feeling-related labels were uncommon overall. Only about one in five self-chosen group names referred to a feeling.

The human groupings also corresponded closely to the structure recovered by the neural network. Pig vocalizations therefore contain genuine, repeatable acoustic information that both people and computers can detect. However, detecting such patterns is not the same as understanding their biological significance. Curious to find out how well you understand pig vocalisations? Take the test at: https://project-oink.org/ 

Exact meaning remains difficult to identify

When participants chose from 18 recording contexts, only 8.0% of answers were correct, compared with 5.6% expected by chance. Judgements of ‘positive’ or ‘negative’ matched the pre-assigned context label in 60.1% of cases when neutral answers were excluded, but this result was driven mainly by highly negative situations such as castration, restraint and fighting.

Once these extreme situations were removed, agreement with the positive and negative labels disappeared. Vocalizations from an enriched pen with straw and from a barren pen with a concrete floor were grouped similarly; the latter were even rated slightly more positively on average. Sound contained structure, but that structure did not reliably reveal the original context or presumed affective category.