Scaling Self-Supervised Speech Models Uncovers Deep Linguistic Relationships: Evidence from the Pacific Cluster

The study presents findings on scaling Self-Supervised Speech Models (S3Ms) from 126 to 4,017 languages, revealing that the larger model (4K) significantly enhances the recovery of linguistic phylogenies, unlike the 1K model which showed flat performance. A notable outcome is the emergence of a Pacific macro-cluster that groups unrelated languages based on shared acoustic features, indicating that large S3Ms can capture complex linguistic relationships and dynamics. This advancement is critical for practitioners in computational linguistics and phylogenetics, as it suggests new methodologies for analyzing language evolution and contact through large-scale speech models.