Ungulate Emotional Valence

Classify emotional valence (positive vs. negative) and species from ungulate contact calls using avex embeddings.

Dataset: Zenodo 14636641 — supplemental data for Machine Learning Algorithms Can Predict Emotional Valence Across Ungulate Vocalizations

Dataset

  • Species: Cow, Pig, Sheep, Goat, Horse, Przewalski’s Horse, Wild Boar (7 species)

  • Size: 3,181 contact call recordings

  • Labels: emotional valence (positive / negative)

  • Source: Zenodo (CC BY 4.0)

Pipeline

download dataset ──► explore + annotate ──► embed (BEATs / EfficientNet)
    ──► UMAP ──► training-free metrics ──► linear probe
    ──► attention probe ──► LOSO cross-species eval ──► speed augmentation

Key results

BEATs and EfficientNet embeddings are compared on two tasks:

  1. Valence classification (positive / negative)

  2. Species identification (7 classes)

Cross-species evaluation (LOSO)

Leave-one-species-out probe: train on 6 species, test on the held-out 7th. Measures how well embeddings transfer to an unseen species.

Speed augmentation

Training set augmented with librosa.effects.time_stretch at 14 rates (0.5× – 2.0×). Tests whether temporal-rate invariance improves cross-species generalisation.