Preregistered Discovery and Replication of Trigger Motifs in Orca Vocalizations

This project investigates whether orca (killer whale) vocalizations exhibit reproducible, non-random sequential structure that can be detected without manual call-type labeling. Using publicly available passive hydrophone recordings, we convert vocalization clips into sequences of discrete acoustic units learned from audio embeddings. We then test two related hypotheses: Combinatorial structure hypothesis: Orca vocalizations contain discrete, reusable acoustic building blocks whose sequences are more structured than appropriately randomized controls, and whose multi-unit motifs recur across separate recordings. Trigger-motif hypothesis: Within these sequences, a subset of rare multi-unit motifs reliably precedes abrupt changes in sequence dynamics (“sequence mode shifts”), suggesting the presence of control-like patterns in vocalization sequences. We do not claim semantic meaning, translation, or human-like grammar; our goal is to quantify and validate sequence-level structure and predictive patterns. To evaluate these hypotheses, we use a Combinatorial Structure Test (CST) consisting of (a) unit stability/distinctiveness checks, (b) non-randomness tests via mutual information and entropy relative to null models (time-shuffle, cross-clip shuffle, random assignment), (c) motif recurrence analysis (n-grams recurring across independent recordings), and (d) negative controls on noise. We additionally perform robustness checks including event-level deduplication to reduce near-duplicate clips, parameter sensitivity analyses, matched-control counterfactual tests (“break-the-trigger”), and Markov-preserving nulls to rule out explanations based solely on bigram statistics. Primary expected outcome: Evidence that unit sequences and motifs show statistically detectable structure beyond chance and beyond simple controls, and that some motifs recur across independent events. Secondary expected outcome: Identification of a set of candidate “trigger motifs” that predict near-future sequence mode shifts more often than matched controls. A key component of this registration is a preregistered blind holdout validation: trigger motifs are discovered using only a training subset (split by recording source/site or session), frozen without modification, and then evaluated on an unseen test subset to reduce researcher degrees of freedom and prevent tuning on the holdout data. All analyses are computational, non-invasive, and use public recordings. This registration covers the preregistered analysis plan, success criteria, and reporting commitments for the discovery and validation steps.